int Drumm2(const Epetra_Map& map, bool verbose)
{
//Simple 2-element problem (element as in "finite-element") from
//Clif Drumm. Two triangular elements, one per processor, as shown
//here:
//
// *----*
// 3|\ 2|
// | \ |
// | 0\1|
// | \|
// *----*
// 0 1
//
//Element 0 on processor 0, element 1 on processor 1.
//Processor 0 will own nodes 0,1,3 and processor 1 will own node 2.
//Each processor will pass a 3x3 element-matrix to Epetra_FECrsMatrix.
//After GlobalAssemble(), the matrix should be as follows:
//
// row 0: 2 1 0 1
//proc 0 row 1: 1 4 1 2
// row 2: 0 1 2 1
//----------------------------------
//proc 1 row 3: 1 2 1 4
//
int numProcs = map.Comm().NumProc();
int localProc = map.Comm().MyPID();
if (numProcs != 2) return(0);
int indexBase = 0, ierr = 0;
double* values = new double[9];
values[0] = 2.0;
values[1] = 1.0;
values[2] = 1.0;
values[3] = 1.0;
values[4] = 2.0;
values[5] = 1.0;
values[6] = 1.0;
values[7] = 1.0;
values[8] = 2.0;
if (localProc == 0) {
int numMyNodes = 3;
int* myNodes = new int[numMyNodes];
myNodes[0] = 0;
myNodes[1] = 1;
myNodes[2] = 3;
Epetra_Map Map(-1, numMyNodes, myNodes, indexBase, map.Comm());
int rowLengths = 3;
Epetra_FECrsMatrix A(Copy, Map, rowLengths);
EPETRA_TEST_ERR( A.InsertGlobalValues(numMyNodes, myNodes,
numMyNodes, myNodes,
values, Epetra_FECrsMatrix::ROW_MAJOR),ierr);
EPETRA_TEST_ERR( A.GlobalAssemble(), ierr );
EPETRA_TEST_ERR( A.GlobalAssemble(), ierr );
if (verbose) {
A.Print(cout);
}
//now let's make sure we can do a matvec with this matrix.
Epetra_Vector x(Map), y(Map);
x.PutScalar(1.0);
EPETRA_TEST_ERR( A.Multiply(false, x, y), ierr);
if (verbose&&localProc==0) {
cout << "y = A*x, x=1.0's"<<endl;
}
if (verbose) {
y.Print(cout);
}
delete [] myNodes;
delete [] values;
}
else {
int numMyNodes = 1;
int* myNodes = new int[numMyNodes];
myNodes[0] = 2;
Epetra_Map Map(-1, numMyNodes, myNodes, indexBase, map.Comm());
int rowLengths = 3;
Epetra_FECrsMatrix A(Copy, Map, rowLengths);
delete [] myNodes;
numMyNodes = 3;
myNodes = new int[numMyNodes];
myNodes[0] = 1;
myNodes[1] = 2;
myNodes[2] = 3;
EPETRA_TEST_ERR( A.InsertGlobalValues(numMyNodes, myNodes,
numMyNodes, myNodes,
values, Epetra_FECrsMatrix::ROW_MAJOR),ierr);
EPETRA_TEST_ERR( A.GlobalAssemble(), ierr );
EPETRA_TEST_ERR( A.GlobalAssemble(), ierr );
if (verbose) {
A.Print(cout);
}
//now let's make sure we can do a matvec with this matrix.
Epetra_Vector x(Map), y(Map);
x.PutScalar(1.0);
EPETRA_TEST_ERR( A.Multiply(false, x, y), ierr);
if (verbose) {
y.Print(cout);
}
delete [] myNodes;
delete [] values;
}
return(0);
}
//=============================================================================
Epetra_Map * Epetra_Map::RemoveEmptyProcesses() const
{
#ifdef HAVE_MPI
const Epetra_MpiComm * MpiComm = dynamic_cast<const Epetra_MpiComm*>(&Comm());
// If the Comm isn't MPI, just treat this as a copy constructor
if(!MpiComm) return new Epetra_Map(*this);
MPI_Comm NewComm,MyMPIComm = MpiComm->Comm();
// Create the new communicator. MPI_Comm_split returns a valid
// communicator on all processes. On processes where color == MPI_UNDEFINED,
// ignore the result. Passing key == 0 tells MPI to order the
// processes in the new communicator by their rank in the old
// communicator.
const int color = (NumMyElements() == 0) ? MPI_UNDEFINED : 1;
// MPI_Comm_split must be called collectively over the original
// communicator. We can't just call it on processes with color
// one, even though we will ignore its result on processes with
// color zero.
int rv = MPI_Comm_split(MyMPIComm,color,0,&NewComm);
if(rv!=MPI_SUCCESS) throw ReportError("Epetra_Map::RemoveEmptyProcesses: MPI_Comm_split failed.",-1);
if(color == MPI_UNDEFINED)
return 0; // We're not in the new map
else {
Epetra_MpiComm * NewEpetraComm = new Epetra_MpiComm(NewComm);
// Use the copy constructor for a new map, but basically because it does nothing useful
Epetra_Map * NewMap = new Epetra_Map(*this);
// Get rid of the old BlockMapData, now make a new one from scratch...
NewMap->CleanupData();
if(GlobalIndicesInt()) {
#ifndef EPETRA_NO_32BIT_GLOBAL_INDICES
NewMap->BlockMapData_ = new Epetra_BlockMapData(NumGlobalElements(),0,IndexBase(),*NewEpetraComm,false);
#endif
}
else {
#ifndef EPETRA_NO_64BIT_GLOBAL_INDICES
NewMap->BlockMapData_ = new Epetra_BlockMapData(NumGlobalElements64(),0,IndexBase64(),*NewEpetraComm,true);
#endif
}
// Now copy all of the relevent bits of BlockMapData...
// NewMap->BlockMapData_->Comm_ = NewEpetraComm;
NewMap->BlockMapData_->LID_ = BlockMapData_->LID_;
#ifndef EPETRA_NO_32BIT_GLOBAL_INDICES
NewMap->BlockMapData_->MyGlobalElements_int_ = BlockMapData_->MyGlobalElements_int_;
#endif
#ifndef EPETRA_NO_64BIT_GLOBAL_INDICES
NewMap->BlockMapData_->MyGlobalElements_LL_ = BlockMapData_->MyGlobalElements_LL_;
#endif
NewMap->BlockMapData_->FirstPointInElementList_ = BlockMapData_->FirstPointInElementList_;
NewMap->BlockMapData_->ElementSizeList_ = BlockMapData_->ElementSizeList_;
NewMap->BlockMapData_->PointToElementList_ = BlockMapData_->PointToElementList_;
NewMap->BlockMapData_->NumGlobalElements_ = BlockMapData_->NumGlobalElements_;
NewMap->BlockMapData_->NumMyElements_ = BlockMapData_->NumMyElements_;
NewMap->BlockMapData_->IndexBase_ = BlockMapData_->IndexBase_;
NewMap->BlockMapData_->ElementSize_ = BlockMapData_->ElementSize_;
NewMap->BlockMapData_->MinMyElementSize_ = BlockMapData_->MinMyElementSize_;
NewMap->BlockMapData_->MaxMyElementSize_ = BlockMapData_->MaxMyElementSize_;
NewMap->BlockMapData_->MinElementSize_ = BlockMapData_->MinElementSize_;
NewMap->BlockMapData_->MaxElementSize_ = BlockMapData_->MaxElementSize_;
NewMap->BlockMapData_->MinAllGID_ = BlockMapData_->MinAllGID_;
NewMap->BlockMapData_->MaxAllGID_ = BlockMapData_->MaxAllGID_;
NewMap->BlockMapData_->MinMyGID_ = BlockMapData_->MinMyGID_;
NewMap->BlockMapData_->MaxMyGID_ = BlockMapData_->MaxMyGID_;
NewMap->BlockMapData_->MinLID_ = BlockMapData_->MinLID_;
NewMap->BlockMapData_->MaxLID_ = BlockMapData_->MaxLID_;
NewMap->BlockMapData_->NumGlobalPoints_ = BlockMapData_->NumGlobalPoints_;
NewMap->BlockMapData_->NumMyPoints_ = BlockMapData_->NumMyPoints_;
NewMap->BlockMapData_->ConstantElementSize_ = BlockMapData_->ConstantElementSize_;
NewMap->BlockMapData_->LinearMap_ = BlockMapData_->LinearMap_;
NewMap->BlockMapData_->DistributedGlobal_ = NewEpetraComm->NumProc()==1 ? false : BlockMapData_->DistributedGlobal_;
NewMap->BlockMapData_->OneToOneIsDetermined_ = BlockMapData_->OneToOneIsDetermined_;
NewMap->BlockMapData_->OneToOne_ = BlockMapData_->OneToOne_;
NewMap->BlockMapData_->GlobalIndicesInt_ = BlockMapData_->GlobalIndicesInt_;
NewMap->BlockMapData_->GlobalIndicesLongLong_ = BlockMapData_->GlobalIndicesLongLong_;
NewMap->BlockMapData_->LastContiguousGID_ = BlockMapData_->LastContiguousGID_;
NewMap->BlockMapData_->LastContiguousGIDLoc_ = BlockMapData_->LastContiguousGIDLoc_;
NewMap->BlockMapData_->LIDHash_ = BlockMapData_->LIDHash_ ? new Epetra_HashTable<int>(*BlockMapData_->LIDHash_) : 0;
// Delay directory construction
NewMap->BlockMapData_->Directory_ = 0;
// Cleanup
delete NewEpetraComm;
return NewMap;
}
#else
// MPI isn't compiled, so just treat this as a copy constructor
return new Epetra_Map(*this);
#endif
}
int main(int argc, char *argv[]) {
using Teuchos::RCP; // reference count pointers
using Teuchos::rcp; //
//
// MPI initialization using Teuchos
//
#ifdef HAVE_MPI
MPI_Init(&argc, &argv);
Epetra_MpiComm comm(MPI_COMM_WORLD);
#else
Epetra_SerialComm comm;
#endif
//
// Parameters
//
GlobalOrdinal numGlobalElements = 256; // problem size
//
// Construct the problem
//
// Construct a Map that puts approximately the same number of equations on each processor
const Epetra_Map map(numGlobalElements, 0, comm);
// Get update list and number of local equations from newly created map.
const size_t numMyElements = map.NumMyElements();
const GlobalOrdinal* myGlobalElements = map.MyGlobalElements();
// Create a CrsMatrix using the map, with a dynamic allocation of 3 entries per row
RCP<Epetra_CrsMatrix> A = rcp(new Epetra_CrsMatrix(Copy, map, 3));
// Add rows one-at-a-time
for (size_t i = 0; i < numMyElements; i++) {
if (myGlobalElements[i] == 0) {
//TODO: should be rewritten in an Epetra style
A->InsertGlobalValues(myGlobalElements[i], 2,
Teuchos::tuple<Scalar> (2.0, -1.0).getRawPtr(),
Teuchos::tuple<GlobalOrdinal>(myGlobalElements[i], myGlobalElements[i] +1).getRawPtr());
}
else if (myGlobalElements[i] == numGlobalElements - 1) {
A->InsertGlobalValues(myGlobalElements[i], 2,
Teuchos::tuple<Scalar> (-1.0, 2.0).getRawPtr(),
Teuchos::tuple<GlobalOrdinal>(myGlobalElements[i] -1, myGlobalElements[i]).getRawPtr());
}
else {
A->InsertGlobalValues(myGlobalElements[i], 3,
Teuchos::tuple<Scalar> (-1.0, 2.0, -1.0).getRawPtr(),
Teuchos::tuple<GlobalOrdinal>(myGlobalElements[i] -1, myGlobalElements[i], myGlobalElements[i] +1).getRawPtr());
}
}
// Complete the fill, ask that storage be reallocated and optimized
A->FillComplete();
//
// Construct a multigrid preconditioner
//
// Turns a Epetra_CrsMatrix into a MueLu::Matrix
RCP<Xpetra::CrsMatrix<SC, LO, GO, NO, LMO> > mueluA_ = rcp(new Xpetra::EpetraCrsMatrix(A)); //TODO: should not be needed
RCP<Xpetra::Matrix <SC, LO, GO, NO, LMO> > mueluA = rcp(new Xpetra::CrsMatrixWrap<SC, LO, GO, NO, LMO>(mueluA_));
// Multigrid Hierarchy
RCP<Hierarchy> H = rcp(new Hierarchy(mueluA));
H->setVerbLevel(Teuchos::VERB_HIGH);
// Multigrid setup phase (using default parameters)
H->Setup();
//
// Define RHS / LHS
//
RCP<Epetra_Vector> X = rcp(new Epetra_Vector(map));
RCP<Epetra_Vector> B = rcp(new Epetra_Vector(map));
X->PutScalar((Scalar) 0.0);
B->SetSeed(846930886); B->Random();
#ifndef HAVE_MUELU_BELOS
//
// Use AMG directly as an iterative solver (not as a preconditionner)
//
int nIts = 9;
// Wrap Epetra Vectors into Xpetra Vectors
RCP<Vector> mueluX = rcp(new Xpetra::EpetraVector(X));
RCP<Vector> mueluB = rcp(new Xpetra::EpetraVector(B));
H->Iterate(*mueluB, nIts, *mueluX);
// Print relative residual norm
ST::magnitudeType residualNorms = Utils::ResidualNorm(*mueluA, *mueluX, *mueluB)[0];
if (comm.MyPID() == 0)
std::cout << "||Residual|| = " << residualNorms << std::endl;
#else // HAVE_MUELU_BELOS
//
// Solve Ax = b using AMG as a preconditioner in Belos
//
// Matrix and Multivector type that will be used with Belos
typedef Epetra_MultiVector MV;
typedef Belos::OperatorT<MV> OP;
// Define Operator and Preconditioner
RCP<OP> belosOp = rcp(new Belos::XpetraOp<SC, LO, GO, NO, LMO>(mueluA)); // Turns a Xpetra::Matrix object into a Belos operator
RCP<OP> belosPrec = rcp(new Belos::MueLuOp<SC, LO, GO, NO, LMO>(H)); // Turns a MueLu::Hierarchy object into a Belos operator
// Construct a Belos LinearProblem object
RCP< Belos::LinearProblem<SC, MV, OP> > belosProblem = rcp(new Belos::LinearProblem<SC, MV, OP>(belosOp, X, B));
belosProblem->setLeftPrec(belosPrec);
bool set = belosProblem->setProblem();
if (set == false) {
std::cout << std::endl << "ERROR: Belos::LinearProblem failed to set up correctly!" << std::endl;
return EXIT_FAILURE;
}
// Belos parameter list
int maxIts = 20;
double tol = 1e-4;
Teuchos::ParameterList belosList;
belosList.set("Maximum Iterations", maxIts); // Maximum number of iterations allowed
belosList.set("Convergence Tolerance", tol); // Relative convergence tolerance requested
belosList.set("Verbosity", Belos::Errors + Belos::Warnings + Belos::TimingDetails + Belos::StatusTestDetails);
// Create an iterative solver manager
RCP< Belos::SolverManager<SC, MV, OP> > solver = rcp(new Belos::BlockCGSolMgr<SC, MV, OP>(belosProblem, rcp(&belosList, false)));
// Perform solve
Belos::ReturnType ret = solver->solve();
// Get the number of iterations for this solve.
std::cout << "Number of iterations performed for this solve: " << solver->getNumIters() << std::endl;
// Compute actual residuals.
int numrhs=1;
bool badRes = false;
std::vector<SC> actual_resids(numrhs);
std::vector<SC> rhs_norm(numrhs);
RCP<Epetra_MultiVector > resid = rcp(new Epetra_MultiVector(map, numrhs));
typedef Belos::OperatorTraits<SC, MV, OP> OPT;
typedef Belos::MultiVecTraits<SC, MV> MVT;
OPT::Apply(*belosOp, *X, *resid);
MVT::MvAddMv(-1.0, *resid, 1.0, *B, *resid);
MVT::MvNorm(*resid, actual_resids);
MVT::MvNorm(*B, rhs_norm);
std::cout<< "---------- Actual Residuals (normalized) ----------"<<std::endl<<std::endl;
for (int i = 0; i < numrhs; i++) {
SC actRes = actual_resids[i]/rhs_norm[i];
std::cout <<"Problem " << i << " : \t" << actRes <<std::endl;
if (actRes > tol) { badRes = true; }
}
// Check convergence
if (ret != Belos::Converged || badRes) {
std::cout << std::endl << "ERROR: Belos did not converge! " << std::endl;
return EXIT_FAILURE;
}
std::cout << std::endl << "SUCCESS: Belos converged!" << std::endl;
#endif // HAVE_MUELU_BELOS
#ifdef HAVE_MPI
MPI_Finalize();
#endif
return EXIT_SUCCESS;
}
Epetra_CrsGraph * BlockUtility::TGenerateBlockGraph(
const Epetra_CrsGraph & BaseGraph,
const Epetra_CrsGraph & LocalBlockGraph,
const Epetra_Comm & GlobalComm )
{
const Epetra_BlockMap & BaseRowMap = BaseGraph.RowMap();
const Epetra_BlockMap & BaseColMap = BaseGraph.ColMap();
int_type ROffset = BlockUtility::TCalculateOffset<int_type>(BaseRowMap);
(void) ROffset; // Silence "unused variable" compiler warning.
int_type COffset = BlockUtility::TCalculateOffset<int_type>(BaseColMap);
//Get Base Global IDs
const Epetra_BlockMap & BlockRowMap = LocalBlockGraph.RowMap();
const Epetra_BlockMap & BlockColMap = LocalBlockGraph.ColMap();
int NumBlockRows = BlockRowMap.NumMyElements();
vector<int_type> RowIndices(NumBlockRows);
BlockRowMap.MyGlobalElements(&RowIndices[0]);
int Size = BaseRowMap.NumMyElements();
Epetra_Map *GlobalRowMap =
GenerateBlockMap(BaseRowMap, BlockRowMap, GlobalComm);
int MaxIndices = BaseGraph.MaxNumIndices();
vector<int_type> Indices(MaxIndices);
Epetra_CrsGraph * GlobalGraph = new Epetra_CrsGraph( Copy,
dynamic_cast<Epetra_BlockMap&>(*GlobalRowMap),
0 );
int NumBlockIndices, NumBaseIndices;
int *BlockIndices, *BaseIndices;
for( int i = 0; i < NumBlockRows; ++i )
{
LocalBlockGraph.ExtractMyRowView(i, NumBlockIndices, BlockIndices);
for( int j = 0; j < Size; ++j )
{
int_type GlobalRow = (int_type) GlobalRowMap->GID64(j+i*Size);
BaseGraph.ExtractMyRowView( j, NumBaseIndices, BaseIndices );
for( int k = 0; k < NumBlockIndices; ++k )
{
int_type ColOffset = (int_type) BlockColMap.GID64(BlockIndices[k]) * COffset;
for( int l = 0; l < NumBaseIndices; ++l )
Indices[l] = (int_type) BaseGraph.GCID64(BaseIndices[l]) + ColOffset;
GlobalGraph->InsertGlobalIndices( GlobalRow, NumBaseIndices, &Indices[0] );
}
}
}
const Epetra_BlockMap & BaseDomainMap = BaseGraph.DomainMap();
const Epetra_BlockMap & BaseRangeMap = BaseGraph.RangeMap();
const Epetra_BlockMap & BlockDomainMap = LocalBlockGraph.DomainMap();
const Epetra_BlockMap & BlockRangeMap = LocalBlockGraph.RangeMap();
Epetra_Map *GlobalDomainMap =
GenerateBlockMap(BaseDomainMap, BlockDomainMap, GlobalComm);
Epetra_Map *GlobalRangeMap =
GenerateBlockMap(BaseRangeMap, BlockRangeMap, GlobalComm);
GlobalGraph->FillComplete(*GlobalDomainMap, *GlobalRangeMap);
delete GlobalDomainMap;
delete GlobalRangeMap;
delete GlobalRowMap;
return GlobalGraph;
}
// FIXME long long
Epetra_Map
Epetra_Util::Create_Root_Map(const Epetra_Map& usermap,
int root)
{
int numProc = usermap.Comm().NumProc();
if (numProc==1) {
Epetra_Map newmap(usermap);
return(newmap);
}
const Epetra_Comm & comm = usermap.Comm();
bool isRoot = usermap.Comm().MyPID()==root;
//if usermap is already completely owned by root then we'll just return a copy of it.
int quickreturn = 0;
int globalquickreturn = 0;
if (isRoot) {
if (usermap.NumMyElements()==usermap.NumGlobalElements64()) quickreturn = 1;
}
else {
if (usermap.NumMyElements()==0) quickreturn = 1;
}
usermap.Comm().MinAll(&quickreturn, &globalquickreturn, 1);
if (globalquickreturn==1) {
Epetra_Map newmap(usermap);
return(newmap);
}
// Linear map: Simple case, just put all GIDs linearly on root processor
if (usermap.LinearMap() && root!=-1) {
int numMyElements = 0;
if (isRoot) numMyElements = usermap.MaxAllGID64()+1; // FIXME long long
Epetra_Map newmap(-1, numMyElements, usermap.IndexBase(), comm);
return(newmap);
}
if (!usermap.UniqueGIDs())
throw usermap.ReportError("usermap must have unique GIDs",-1);
// General map
// Build IntVector of the GIDs, then ship them to root processor
int numMyElements = usermap.NumMyElements();
Epetra_Map allGidsMap(-1, numMyElements, 0, comm);
Epetra_IntVector allGids(allGidsMap);
for (int i=0; i<numMyElements; i++) allGids[i] = usermap.GID64(i);
int numGlobalElements = usermap.NumGlobalElements64();
if (root!=-1) {
int n1 = 0; if (isRoot) n1 = numGlobalElements;
Epetra_Map allGidsOnRootMap(-1, n1, 0, comm);
Epetra_Import importer(allGidsOnRootMap, allGidsMap);
Epetra_IntVector allGidsOnRoot(allGidsOnRootMap);
allGidsOnRoot.Import(allGids, importer, Insert);
Epetra_Map rootMap(-1, allGidsOnRoot.MyLength(), allGidsOnRoot.Values(), usermap.IndexBase(), comm);
return(rootMap);
}
else {
int n1 = numGlobalElements;
Epetra_LocalMap allGidsOnRootMap(n1, 0, comm);
Epetra_Import importer(allGidsOnRootMap, allGidsMap);
Epetra_IntVector allGidsOnRoot(allGidsOnRootMap);
allGidsOnRoot.Import(allGids, importer, Insert);
Epetra_Map rootMap(-1, allGidsOnRoot.MyLength(), allGidsOnRoot.Values(), usermap.IndexBase(), comm);
return(rootMap);
}
}
int TLowCommunicationMakeColMapAndReindex(int N, const int * rowptr, int * colind_LID, const int_type *colind_GID, const Epetra_Map& domainMap, const int * owningPIDs, bool SortGhostsAssociatedWithEachProcessor, std::vector<int>& RemotePIDs, MapType1 & NewColMap)
{
int i,j;
// Sanity checks
bool UseLL;
if(domainMap.GlobalIndicesLongLong()) UseLL=true;
else if(domainMap.GlobalIndicesInt()) UseLL=false;
else throw std::runtime_error("LowCommunicationMakeColMapAndReindex: cannot detect int type.");
// Scan all column indices and sort into two groups:
// Local: those whose GID matches a GID of the domain map on this processor and
// Remote: All others.
int numDomainElements = domainMap.NumMyElements();
bool * LocalGIDs = 0;
if (numDomainElements>0) LocalGIDs = new bool[numDomainElements];
for (i=0; i<numDomainElements; i++) LocalGIDs[i] = false; // Assume domain GIDs are not local
bool DoSizes = !domainMap.ConstantElementSize(); // If not constant element size, then error
if(DoSizes) throw std::runtime_error("LowCommunicationMakeColMapAndReindex: cannot handle non-constant sized domainMap.");
// In principle it is good to have RemoteGIDs and RemotGIDList be as long as the number of remote GIDs
// on this processor, but this would require two passes through the column IDs, so we make it the max of 100
// and the number of block rows.
const int numMyBlockRows = N;
int hashsize = numMyBlockRows; if (hashsize < 100) hashsize = 100;
Epetra_HashTable<int_type> RemoteGIDs(hashsize);
std::vector<int_type> RemoteGIDList; RemoteGIDList.reserve(hashsize);
std::vector<int> PIDList; PIDList.reserve(hashsize);
// Here we start using the *int* colind array. If int_type==int this clobbers the GIDs, if
// int_type==long long, then this is the first use of the colind array.
// For *local* GID's set colind with with their LID in the domainMap. For *remote* GIDs,
// we set colind with (numDomainElements+NumRemoteColGIDs) before the increment of
// the remote count. These numberings will be separate because no local LID is greater
// than numDomainElements.
int NumLocalColGIDs = 0;
int NumRemoteColGIDs = 0;
for(i = 0; i < numMyBlockRows; i++) {
for(j = rowptr[i]; j < rowptr[i+1]; j++) {
int_type GID = colind_GID[j];
// Check if GID matches a row GID
int LID = domainMap.LID(GID);
if(LID != -1) {
bool alreadyFound = LocalGIDs[LID];
if (!alreadyFound) {
LocalGIDs[LID] = true; // There is a column in the graph associated with this domain map GID
NumLocalColGIDs++;
}
colind_LID[j] = LID;
}
else {
int_type hash_value=RemoteGIDs.Get(GID);
if(hash_value == -1) { // This means its a new remote GID
int PID = owningPIDs[j];
if(PID==-1) throw std::runtime_error("LowCommunicationMakeColMapAndReindex: Cannot figure out if PID is owned.");
colind_LID[j] = numDomainElements + NumRemoteColGIDs;
RemoteGIDs.Add(GID, NumRemoteColGIDs);
RemoteGIDList.push_back(GID);
PIDList.push_back(PID);
NumRemoteColGIDs++;
}
else
colind_LID[j] = numDomainElements + hash_value;
}
}
}
// Possible short-circuit: If all domain map GIDs are present as column indices, then set ColMap=domainMap and quit
if (domainMap.Comm().NumProc()==1) {
if (NumRemoteColGIDs!=0) {
throw std::runtime_error("Some column IDs are not in domainMap. If matrix is rectangular, you must pass in a domainMap");
// Sanity test: When one processor,there can be no remoteGIDs
}
if (NumLocalColGIDs==numDomainElements) {
if (LocalGIDs!=0) delete [] LocalGIDs;
// In this case, we just use the domainMap's indices, which is, not coincidently, what we clobbered colind with up above anyway.
// No further reindexing is needed.
NewColMap = domainMap;
return 0;
}
}
// Now build integer array containing column GIDs
// Build back end, containing remote GIDs, first
int numMyBlockCols = NumLocalColGIDs + NumRemoteColGIDs;
std::vector<int_type> ColIndices;
int_type * RemoteColIndices=0;
if(numMyBlockCols > 0) {
ColIndices.resize(numMyBlockCols);
if(NumLocalColGIDs!=numMyBlockCols) RemoteColIndices = &ColIndices[NumLocalColGIDs]; // Points to back end of ColIndices
else RemoteColIndices=0;
}
for(i = 0; i < NumRemoteColGIDs; i++)
RemoteColIndices[i] = RemoteGIDList[i];
// Build permute array for *remote* reindexing.
std::vector<int> RemotePermuteIDs(NumRemoteColGIDs);
for(i=0; i<NumRemoteColGIDs; i++) RemotePermuteIDs[i]=i;
// Sort External column indices so that all columns coming from a given remote processor are contiguous
int NumListsInt=0;
int NumListsLL =0;
int * IntSortLists[2];
long long * LLSortLists[2];
int * RemotePermuteIDs_ptr = RemotePermuteIDs.size() ? &RemotePermuteIDs[0] : 0;
if(!UseLL) {
// int version
IntSortLists[0] = (int*) RemoteColIndices;
IntSortLists[1] = RemotePermuteIDs_ptr;
NumListsInt=2;
}
else {
//LL version
LLSortLists[0] = (long long*) RemoteColIndices;
IntSortLists[0] = RemotePermuteIDs_ptr;
NumListsInt = NumListsLL = 1;
}
int * PIDList_ptr = PIDList.size() ? &PIDList[0] : 0;
Epetra_Util::Sort(true, NumRemoteColGIDs, PIDList_ptr, 0, 0, NumListsInt, IntSortLists,NumListsLL,LLSortLists);
// Stash the RemotePIDs
PIDList.resize(NumRemoteColGIDs);
RemotePIDs = PIDList;
if (SortGhostsAssociatedWithEachProcessor) {
// Sort external column indices so that columns from a given remote processor are not only contiguous
// but also in ascending order. NOTE: I don't know if the number of externals associated
// with a given remote processor is known at this point ... so I count them here.
// NTS: Only sort the RemoteColIndices this time...
int StartCurrent, StartNext;
StartCurrent = 0; StartNext = 1;
while ( StartNext < NumRemoteColGIDs ) {
if (PIDList[StartNext]==PIDList[StartNext-1]) StartNext++;
else {
IntSortLists[0] = &RemotePermuteIDs[StartCurrent];
Epetra_Util::Sort(true,StartNext-StartCurrent, &(RemoteColIndices[StartCurrent]),0,0,1,IntSortLists,0,0);
StartCurrent = StartNext; StartNext++;
}
}
IntSortLists[0] = &RemotePermuteIDs[StartCurrent];
Epetra_Util::Sort(true, StartNext-StartCurrent, &(RemoteColIndices[StartCurrent]), 0, 0, 1,IntSortLists,0,0);
}
// Reverse the permutation to get the information we actually care about
std::vector<int> ReverseRemotePermuteIDs(NumRemoteColGIDs);
for(i=0; i<NumRemoteColGIDs; i++) ReverseRemotePermuteIDs[RemotePermuteIDs[i]]=i;
// Build permute array for *local* reindexing.
bool use_local_permute=false;
std::vector<int> LocalPermuteIDs(numDomainElements);
// Now fill front end. Two cases:
// (1) If the number of Local column GIDs is the same as the number of Local domain GIDs, we
// can simply read the domain GIDs into the front part of ColIndices, otherwise
// (2) We step through the GIDs of the domainMap, checking to see if each domain GID is a column GID.
// we want to do this to maintain a consistent ordering of GIDs between the columns and the domain.
if(NumLocalColGIDs == domainMap.NumMyElements()) {
if(NumLocalColGIDs > 0) {
domainMap.MyGlobalElements(&ColIndices[0]); // Load Global Indices into first numMyBlockCols elements column GID list
}
}
else {
int_type* MyGlobalElements = 0;
domainMap.MyGlobalElementsPtr(MyGlobalElements);
int* ElementSizeList = 0;
if(DoSizes)
ElementSizeList = domainMap.ElementSizeList();
int NumLocalAgain = 0;
use_local_permute = true;
for(i = 0; i < numDomainElements; i++) {
if(LocalGIDs[i]) {
LocalPermuteIDs[i] = NumLocalAgain;
ColIndices[NumLocalAgain++] = MyGlobalElements[i];
}
}
assert(NumLocalAgain==NumLocalColGIDs); // Sanity test
}
// Done with this array
if (LocalGIDs!=0) delete [] LocalGIDs;
// Make Column map with same element sizes as Domain map
int_type * ColIndices_ptr = ColIndices.size() ? &ColIndices[0] : 0;
MapType2 temp((int_type)(-1), numMyBlockCols, ColIndices_ptr, (int_type)domainMap.IndexBase64(), domainMap.Comm());
NewColMap = temp;
// Low-cost reindex of the matrix
for(i=0; i<numMyBlockRows; i++){
for(j=rowptr[i]; j<rowptr[i+1]; j++){
int ID=colind_LID[j];
if(ID < numDomainElements){
if(use_local_permute) colind_LID[j] = LocalPermuteIDs[colind_LID[j]];
// In the case where use_local_permute==false, we just copy the DomainMap's ordering, which it so happens
// is what we put in colind to begin with.
}
else
colind_LID[j] = NumLocalColGIDs + ReverseRemotePermuteIDs[colind_LID[j]-numDomainElements];
}
}
return 0;
}
int powerMethodTests(Epetra_RowMatrix & A, Epetra_RowMatrix & JadA, Epetra_Map & Map,
Epetra_Vector & q, Epetra_Vector & z, Epetra_Vector & resid, bool verbose) {
// variable needed for iteration
double lambda = 0.0;
// int niters = 10000;
int niters = 300;
double tolerance = 1.0e-2;
int ierr = 0;
/////////////////////////////////////////////////////////////////////////////////////////////////
// Iterate
Epetra_Time timer(Map.Comm());
double startTime = timer.ElapsedTime();
EPETRA_TEST_ERR(power_method(false, A, q, z, resid, &lambda, niters, tolerance, verbose),ierr);
double elapsed_time = timer.ElapsedTime() - startTime;
double total_flops = q.Flops();
double MFLOPs = total_flops/elapsed_time/1000000.0;
double lambdaref = lambda;
double flopsref = total_flops;
if (verbose)
cout << "\n\nTotal MFLOPs for reference first solve = " << MFLOPs << endl
<< "Total FLOPS = " <<total_flops <<endl<<endl;
lambda = 0.0;
startTime = timer.ElapsedTime();
EPETRA_TEST_ERR(power_method(false, JadA, q, z, resid, &lambda, niters, tolerance, verbose),ierr);
elapsed_time = timer.ElapsedTime() - startTime;
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose)
cout << "\n\nTotal MFLOPs for candidate first solve = " << MFLOPs << endl
<< "Total FLOPS = " <<total_flops <<endl<<endl;
EPETRA_TEST_ERR(checkValues(lambda,lambdaref," No-transpose Power Method result", verbose),ierr);
EPETRA_TEST_ERR(checkValues(total_flops,flopsref," No-transpose Power Method flop count", verbose),ierr);
/////////////////////////////////////////////////////////////////////////////////////////////////
// Solve transpose problem
if (verbose) cout << "\n\nUsing transpose of matrix and solving again (should give same result).\n\n"
<< endl;
// Iterate
lambda = 0.0;
startTime = timer.ElapsedTime();
EPETRA_TEST_ERR(power_method(true, A, q, z, resid, &lambda, niters, tolerance, verbose),ierr);
elapsed_time = timer.ElapsedTime() - startTime;
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
lambdaref = lambda;
flopsref = total_flops;
if (verbose)
cout << "\n\nTotal MFLOPs for reference transpose solve = " << MFLOPs << endl
<< "Total FLOPS = " <<total_flops <<endl<<endl;
lambda = 0.0;
startTime = timer.ElapsedTime();
EPETRA_TEST_ERR(power_method(true, JadA, q, z, resid, &lambda, niters, tolerance, verbose),ierr);
elapsed_time = timer.ElapsedTime() - startTime;
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose)
cout << "\n\nTotal MFLOPs for candidate transpose solve = " << MFLOPs << endl
<< "Total FLOPS = " <<total_flops <<endl<<endl;
EPETRA_TEST_ERR(checkValues(lambda,lambdaref,"Transpose Power Method result", verbose),ierr);
EPETRA_TEST_ERR(checkValues(total_flops,flopsref,"Transpose Power Method flop count", verbose),ierr);
EPETRA_TEST_ERR(check(A, JadA, verbose),ierr);
return(0);
}
Epetra_CrsMatrix *
Laplace2D::CreateLaplacian( const int nx, const int ny, const Epetra_Comm * Comm)
{
int NumGlobalElements = nx * ny;
// create a map
Epetra_Map * Map = new Epetra_Map(NumGlobalElements,0,*Comm);
// local number of rows
int NumMyElements = Map->NumMyElements();
// get update list
int * MyGlobalElements = Map->MyGlobalElements();
double hx = 1.0/(nx-1);
double hy = 1.0/(ny-1);
double off_left = -1.0/(hx*hx);
double off_right = -1.0/(hx*hx);
double off_lower = -1.0/(hy*hy);
double off_upper = -1.0/(hy*hy);
double diag = 2.0/(hx*hx) + 2.0/(hy*hy);
int left, right, lower, upper;
// a bit overestimated the nonzero per row
Epetra_CrsMatrix * A = new Epetra_CrsMatrix(Copy,*Map,5);
// Add rows one-at-a-time
double * Values = new double[4];
int * Indices = new int[4];
for( int i = 0; i < NumMyElements; ++i )
{
int NumEntries=0;
get_myNeighbours( MyGlobalElements[i], nx, ny, left, right, lower, upper );
if( left != -1 )
{
Indices[NumEntries] = left;
Values[NumEntries] = off_left;
++NumEntries;
}
if( right != -1 )
{
Indices[NumEntries] = right;
Values[NumEntries] = off_right;
++NumEntries;
}
if( lower != -1 )
{
Indices[NumEntries] = lower;
Values[NumEntries] = off_lower;
++NumEntries;
}
if( upper != -1 )
{
Indices[NumEntries] = upper;
Values[NumEntries] = off_upper;
++NumEntries;
}
// put the off-diagonal entries
A->InsertGlobalValues(MyGlobalElements[i], NumEntries, Values, Indices);
// Put in the diagonal entry
A->InsertGlobalValues(MyGlobalElements[i], 1, &diag, MyGlobalElements+i);
}
// put matrix in local ordering
A->FillComplete();
delete [] Indices;
delete [] Values;
delete Map;
return A;
} /* createJacobian */
void
LOCA::Epetra::AugmentedOp::buildExtendedMap(const Epetra_BlockMap& uMap,
Epetra_Map*& eMapPtr,
bool buildImporter,
bool haveParam)
{
Epetra_BlockMap& nonconstUnderlyingMap = const_cast<Epetra_BlockMap&>(uMap);
// Convert underlying map to point map if necessary
Epetra_Map* uPointMapPtr =
dynamic_cast<Epetra_Map*>(&nonconstUnderlyingMap);
bool allocatedPointMap = false;
if (uPointMapPtr == NULL) {
allocatedPointMap = true;
blockMap2PointMap(uMap, uPointMapPtr);
}
int max_gid = uPointMapPtr->MaxAllGID();
int num_global_elements = uPointMapPtr->NumGlobalElements();
int num_my_elements = uPointMapPtr->NumMyElements();
int *global_elements = uPointMapPtr->MyGlobalElements();
const Epetra_Comm& comm = uPointMapPtr->Comm();
int index_base = uPointMapPtr->IndexBase();
int ext_num_global_elements;
int ext_num_my_elements;
int *ext_global_elements;
// Compute number of extended global elements
if (buildImporter)
ext_num_global_elements =
num_global_elements + numConstraints*comm.NumProc();
else
ext_num_global_elements = num_global_elements + numConstraints;
// Compute number of extended local elements
if (buildImporter || haveParam)
ext_num_my_elements = num_my_elements + numConstraints;
else
ext_num_my_elements = num_my_elements;
// Allocate extended global elements array
ext_global_elements = new int[ext_num_my_elements];
// Set extended global elements
for (int i=0; i<num_my_elements; i++) {
ext_global_elements[i] = global_elements[i];
}
if (buildImporter || haveParam)
for (int i=0; i<numConstraints; i++)
ext_global_elements[num_my_elements+i] = max_gid + 1 + i;
// Create extended point map
eMapPtr = new Epetra_Map(ext_num_global_elements, ext_num_my_elements,
ext_global_elements, index_base, comm);
// Free global elements array
delete [] ext_global_elements;
if (allocatedPointMap)
delete uPointMapPtr;
}
// ============================================================================
void EpetraExt::XMLWriter::
Write(const std::string& Label, const Epetra_Map& Map)
{
TEUCHOS_TEST_FOR_EXCEPTION(IsOpen_ == false, std::logic_error,
"No file has been opened");
long long NumGlobalElements = Map.NumGlobalElements64();
const int* MyGlobalElements_int = 0;
const long long* MyGlobalElements_LL = 0;
Map.MyGlobalElements(MyGlobalElements_int, MyGlobalElements_LL);
if(!MyGlobalElements_int || !MyGlobalElements_LL)
throw "EpetraExt::XMLWriter::Write: ERROR, GlobalIndices type unknown.";
if (Comm_.MyPID() == 0)
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << "<Map Label=\"" << Label
<< "\" NumElements=\"" << NumGlobalElements << '"'
<< " IndexBase=\"" << Map.IndexBase64() << '"'
<< " NumProc=\"" << Comm_.NumProc() << '"';
of.close();
}
for (int iproc = 0; iproc < Comm_.NumProc(); ++iproc)
{
if (iproc == Comm_.MyPID())
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << " ElementsOnProc" << iproc << "=\"" << Map.NumMyElements() << '"';
of.close();
}
Comm_.Barrier();
}
if (Comm_.MyPID() == 0)
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << '>' << std::endl;
of.close();
}
for (int iproc = 0; iproc < Comm_.NumProc(); iproc++)
{
if (iproc == Comm_.MyPID())
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << "<Proc ID=\"" << Comm_.MyPID() << "\">" << std::endl;
if(MyGlobalElements_int)
{
for (int i = 0; i < Map.NumMyElements(); ++i)
{
of << MyGlobalElements_int[i] << std::endl;
}
}
else
{
for (int i = 0; i < Map.NumMyElements(); ++i)
{
of << MyGlobalElements_LL[i] << std::endl;
}
}
of << "</Proc>" << std::endl;
of.close();
}
Comm_.Barrier();
}
if (Comm_.MyPID() == 0)
{
std::ofstream of(FileName_.c_str(), std::ios::app);
of << "</Map>" << std::endl;
of.close();
}
}
int main(int argc, char *argv[])
{
int ierr = 0;
double elapsed_time;
double total_flops;
double MFLOPs;
#ifdef EPETRA_MPI
// Initialize MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm comm( MPI_COMM_WORLD );
#else
Epetra_SerialComm comm;
#endif
bool verbose = false;
bool summary = false;
// Check if we should print verbose results to standard out
if (argc>6) if (argv[6][0]=='-' && argv[6][1]=='v') verbose = true;
// Check if we should print verbose results to standard out
if (argc>6) if (argv[6][0]=='-' && argv[6][1]=='s') summary = true;
if(argc < 6) {
cerr << "Usage: " << argv[0]
<< " NumNodesX NumNodesY NumProcX NumProcY NumPoints [-v|-s]" << endl
<< "where:" << endl
<< "NumNodesX - Number of mesh nodes in X direction per processor" << endl
<< "NumNodesY - Number of mesh nodes in Y direction per processor" << endl
<< "NumProcX - Number of processors to use in X direction" << endl
<< "NumProcY - Number of processors to use in Y direction" << endl
<< "NumPoints - Number of points to use in stencil (5, 9 or 25 only)" << endl
<< "-v|-s - (Optional) Run in verbose mode if -v present or summary mode if -s present" << endl
<< " NOTES: NumProcX*NumProcY must equal the number of processors used to run the problem." << endl << endl
<< " Serial example:" << endl
<< argv[0] << " 16 12 1 1 25 -v" << endl
<< " Run this program in verbose mode on 1 processor using a 16 X 12 grid with a 25 point stencil."<< endl <<endl
<< " MPI example:" << endl
<< "mpirun -np 32 " << argv[0] << " 10 12 4 8 9 -v" << endl
<< " Run this program in verbose mode on 32 processors putting a 10 X 12 subgrid on each processor using 4 processors "<< endl
<< " in the X direction and 8 in the Y direction. Total grid size is 40 points in X and 96 in Y with a 9 point stencil."<< endl
<< endl;
return(1);
}
//char tmp;
//if (comm.MyPID()==0) cout << "Press any key to continue..."<< endl;
//if (comm.MyPID()==0) cin >> tmp;
//comm.Barrier();
comm.SetTracebackMode(0); // This should shut down any error traceback reporting
if (verbose && comm.MyPID()==0)
cout << Epetra_Version() << endl << endl;
if (summary && comm.MyPID()==0) {
if (comm.NumProc()==1)
cout << Epetra_Version() << endl << endl;
else
cout << endl << endl; // Print two blank line to keep output columns lined up
}
if (verbose) cout << comm <<endl;
// Redefine verbose to only print on PE 0
if (verbose && comm.MyPID()!=0) verbose = false;
if (summary && comm.MyPID()!=0) summary = false;
int numNodesX = atoi(argv[1]);
int numNodesY = atoi(argv[2]);
int numProcsX = atoi(argv[3]);
int numProcsY = atoi(argv[4]);
int numPoints = atoi(argv[5]);
if (verbose || (summary && comm.NumProc()==1)) {
cout << " Number of local nodes in X direction = " << numNodesX << endl
<< " Number of local nodes in Y direction = " << numNodesY << endl
<< " Number of global nodes in X direction = " << numNodesX*numProcsX << endl
<< " Number of global nodes in Y direction = " << numNodesY*numProcsY << endl
<< " Number of local nonzero entries = " << numNodesX*numNodesY*numPoints << endl
<< " Number of global nonzero entries = " << numNodesX*numNodesY*numPoints*numProcsX*numProcsY << endl
<< " Number of Processors in X direction = " << numProcsX << endl
<< " Number of Processors in Y direction = " << numProcsY << endl
<< " Number of Points in stencil = " << numPoints << endl << endl;
}
// Print blank line to keep output columns lined up
if (summary && comm.NumProc()>1)
cout << endl << endl << endl << endl << endl << endl << endl << endl<< endl << endl;
if (numProcsX*numProcsY!=comm.NumProc()) {
cerr << "Number of processors = " << comm.NumProc() << endl
<< " is not the product of " << numProcsX << " and " << numProcsY << endl << endl;
return(1);
}
if (numPoints!=5 && numPoints!=9 && numPoints!=25) {
cerr << "Number of points specified = " << numPoints << endl
<< " is not 5, 9, 25" << endl << endl;
return(1);
}
if (numNodesX*numNodesY<=0) {
cerr << "Product of number of nodes is <= zero" << endl << endl;
return(1);
}
Epetra_IntSerialDenseVector Xoff, XLoff, XUoff;
Epetra_IntSerialDenseVector Yoff, YLoff, YUoff;
if (numPoints==5) {
// Generate a 5-point 2D Finite Difference matrix
Xoff.Size(5);
Yoff.Size(5);
Xoff[0] = -1; Xoff[1] = 1; Xoff[2] = 0; Xoff[3] = 0; Xoff[4] = 0;
Yoff[0] = 0; Yoff[1] = 0; Yoff[2] = 0; Yoff[3] = -1; Yoff[4] = 1;
// Generate a 2-point 2D Lower triangular Finite Difference matrix
XLoff.Size(2);
YLoff.Size(2);
XLoff[0] = -1; XLoff[1] = 0;
YLoff[0] = 0; YLoff[1] = -1;
// Generate a 3-point 2D upper triangular Finite Difference matrix
XUoff.Size(3);
YUoff.Size(3);
XUoff[0] = 0; XUoff[1] = 1; XUoff[2] = 0;
YUoff[0] = 0; YUoff[1] = 0; YUoff[2] = 1;
}
else if (numPoints==9) {
// Generate a 9-point 2D Finite Difference matrix
Xoff.Size(9);
Yoff.Size(9);
Xoff[0] = -1; Xoff[1] = 0; Xoff[2] = 1;
Yoff[0] = -1; Yoff[1] = -1; Yoff[2] = -1;
Xoff[3] = -1; Xoff[4] = 0; Xoff[5] = 1;
Yoff[3] = 0; Yoff[4] = 0; Yoff[5] = 0;
Xoff[6] = -1; Xoff[7] = 0; Xoff[8] = 1;
Yoff[6] = 1; Yoff[7] = 1; Yoff[8] = 1;
// Generate a 5-point lower triangular 2D Finite Difference matrix
XLoff.Size(5);
YLoff.Size(5);
XLoff[0] = -1; XLoff[1] = 0; Xoff[2] = 1;
YLoff[0] = -1; YLoff[1] = -1; Yoff[2] = -1;
XLoff[3] = -1; XLoff[4] = 0;
YLoff[3] = 0; YLoff[4] = 0;
// Generate a 4-point upper triangular 2D Finite Difference matrix
XUoff.Size(4);
YUoff.Size(4);
XUoff[0] = 1;
YUoff[0] = 0;
XUoff[1] = -1; XUoff[2] = 0; XUoff[3] = 1;
YUoff[1] = 1; YUoff[2] = 1; YUoff[3] = 1;
}
else {
// Generate a 25-point 2D Finite Difference matrix
Xoff.Size(25);
Yoff.Size(25);
int xi = 0, yi = 0;
int xo = -2, yo = -2;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
xo = -2, yo++;
Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++; Xoff[xi++] = xo++;
Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ; Yoff[yi++] = yo ;
// Generate a 13-point lower triangular 2D Finite Difference matrix
XLoff.Size(13);
YLoff.Size(13);
xi = 0, yi = 0;
xo = -2, yo = -2;
XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++;
YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ;
xo = -2, yo++;
XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++;
YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ;
xo = -2, yo++;
XLoff[xi++] = xo++; XLoff[xi++] = xo++; XLoff[xi++] = xo++;
YLoff[yi++] = yo ; YLoff[yi++] = yo ; YLoff[yi++] = yo ;
// Generate a 13-point upper triangular 2D Finite Difference matrix
XUoff.Size(13);
YUoff.Size(13);
xi = 0, yi = 0;
xo = 0, yo = 0;
XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++;
YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ;
xo = -2, yo++;
XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++;
YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ;
xo = -2, yo++;
XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++; XUoff[xi++] = xo++;
YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ; YUoff[yi++] = yo ;
}
Epetra_Map * map;
Epetra_Map * mapL;
Epetra_Map * mapU;
Epetra_CrsMatrix * A;
Epetra_CrsMatrix * L;
Epetra_CrsMatrix * U;
Epetra_MultiVector * b;
Epetra_MultiVector * bt;
Epetra_MultiVector * xexact;
Epetra_MultiVector * bL;
Epetra_MultiVector * btL;
Epetra_MultiVector * xexactL;
Epetra_MultiVector * bU;
Epetra_MultiVector * btU;
Epetra_MultiVector * xexactU;
Epetra_SerialDenseVector resvec(0);
//Timings
Epetra_Flops flopcounter;
Epetra_Time timer(comm);
#ifdef EPETRA_VERY_SHORT_PERFTEST
int jstop = 1;
#elif EPETRA_SHORT_PERFTEST
int jstop = 1;
#else
int jstop = 2;
#endif
for (int j=0; j<jstop; j++) {
for (int k=1; k<17; k++) {
#ifdef EPETRA_VERY_SHORT_PERFTEST
if (k<3 || (k%4==0 && k<9)) {
#elif EPETRA_SHORT_PERFTEST
if (k<6 || k%4==0) {
#else
if (k<7 || k%2==0) {
#endif
int nrhs=k;
if (verbose) cout << "\n*************** Results for " << nrhs << " RHS with ";
bool StaticProfile = (j!=0);
if (verbose)
if (StaticProfile) cout << " static profile\n";
else cout << " dynamic profile\n";
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, numPoints,
Xoff.Values(), Yoff.Values(), nrhs, comm, verbose, summary,
map, A, b, bt, xexact, StaticProfile, false);
#ifdef EPETRA_HAVE_JADMATRIX
timer.ResetStartTime();
Epetra_JadMatrix JA(*A);
elapsed_time = timer.ElapsedTime();
if (verbose) cout << "Time to create Jagged diagonal matrix = " << elapsed_time << endl;
//cout << "A = " << *A << endl;
//cout << "JA = " << JA << endl;
runJadMatrixTests(&JA, b, bt, xexact, StaticProfile, verbose, summary);
#endif
runMatrixTests(A, b, bt, xexact, StaticProfile, verbose, summary);
delete A;
delete b;
delete bt;
delete xexact;
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, XLoff.Length(),
XLoff.Values(), YLoff.Values(), nrhs, comm, verbose, summary,
mapL, L, bL, btL, xexactL, StaticProfile, true);
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, XUoff.Length(),
XUoff.Values(), YUoff.Values(), nrhs, comm, verbose, summary,
mapU, U, bU, btU, xexactU, StaticProfile, true);
runLUMatrixTests(L, bL, btL, xexactL, U, bU, btU, xexactU, StaticProfile, verbose, summary);
delete L;
delete bL;
delete btL;
delete xexactL;
delete mapL;
delete U;
delete bU;
delete btU;
delete xexactU;
delete mapU;
Epetra_MultiVector q(*map, nrhs);
Epetra_MultiVector z(q);
Epetra_MultiVector r(q);
delete map;
q.SetFlopCounter(flopcounter);
z.SetFlopCounter(q);
r.SetFlopCounter(q);
resvec.Resize(nrhs);
flopcounter.ResetFlops();
timer.ResetStartTime();
//10 norms
for( int i = 0; i < 10; ++i )
q.Norm2( resvec.Values() );
elapsed_time = timer.ElapsedTime();
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "\nTotal MFLOPs for 10 Norm2's= " << MFLOPs << endl;
if (summary) {
if (comm.NumProc()==1) cout << "Norm2" << '\t';
cout << MFLOPs << endl;
}
flopcounter.ResetFlops();
timer.ResetStartTime();
//10 dot's
for( int i = 0; i < 10; ++i )
q.Dot(z, resvec.Values());
elapsed_time = timer.ElapsedTime();
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Total MFLOPs for 10 Dot's = " << MFLOPs << endl;
if (summary) {
if (comm.NumProc()==1) cout << "DotProd" << '\t';
cout << MFLOPs << endl;
}
flopcounter.ResetFlops();
timer.ResetStartTime();
//10 dot's
for( int i = 0; i < 10; ++i )
q.Update(1.0, z, 1.0, r, 0.0);
elapsed_time = timer.ElapsedTime();
total_flops = q.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Total MFLOPs for 10 Updates= " << MFLOPs << endl;
if (summary) {
if (comm.NumProc()==1) cout << "Update" << '\t';
cout << MFLOPs << endl;
}
}
}
}
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return ierr ;
}
// Constructs a 2D PDE finite difference matrix using the list of x and y offsets.
//
// nx (In) - number of grid points in x direction
// ny (In) - number of grid points in y direction
// The total number of equations will be nx*ny ordered such that the x direction changes
// most rapidly:
// First equation is at point (0,0)
// Second at (1,0)
// ...
// nx equation at (nx-1,0)
// nx+1st equation at (0,1)
// numPoints (In) - number of points in finite difference stencil
// xoff (In) - stencil offsets in x direction (of length numPoints)
// yoff (In) - stencil offsets in y direction (of length numPoints)
// A standard 5-point finite difference stencil would be described as:
// numPoints = 5
// xoff = [-1, 1, 0, 0, 0]
// yoff = [ 0, 0, 0, -1, 1]
// nrhs - Number of rhs to generate. (First interface produces vectors, so nrhs is not needed
// comm (In) - an Epetra_Comm object describing the parallel machine (numProcs and my proc ID)
// map (Out) - Epetra_Map describing distribution of matrix and vectors/multivectors
// A (Out) - Epetra_CrsMatrix constructed for nx by ny grid using prescribed stencil
// Off-diagonal values are random between 0 and 1. If diagonal is part of stencil,
// diagonal will be slightly diag dominant.
// b (Out) - Generated RHS. Values satisfy b = A*xexact
// bt (Out) - Generated RHS. Values satisfy b = A'*xexact
// xexact (Out) - Generated exact solution to Ax = b and b' = A'xexact
// Note: Caller of this function is responsible for deleting all output objects.
void GenerateCrsProblem(int numNodesX, int numNodesY, int numProcsX, int numProcsY, int numPoints,
int * xoff, int * yoff,
const Epetra_Comm &comm, bool verbose, bool summary,
Epetra_Map *& map,
Epetra_CrsMatrix *& A,
Epetra_Vector *& b,
Epetra_Vector *& bt,
Epetra_Vector *&xexact, bool StaticProfile, bool MakeLocalOnly) {
Epetra_MultiVector * b1, * bt1, * xexact1;
GenerateCrsProblem(numNodesX, numNodesY, numProcsX, numProcsY, numPoints,
xoff, yoff, 1, comm, verbose, summary,
map, A, b1, bt1, xexact1, StaticProfile, MakeLocalOnly);
b = dynamic_cast<Epetra_Vector *>(b1);
bt = dynamic_cast<Epetra_Vector *>(bt1);
xexact = dynamic_cast<Epetra_Vector *>(xexact1);
return;
}
void GenerateCrsProblem(int numNodesX, int numNodesY, int numProcsX, int numProcsY, int numPoints,
int * xoff, int * yoff, int nrhs,
const Epetra_Comm &comm, bool verbose, bool summary,
Epetra_Map *& map,
Epetra_CrsMatrix *& A,
Epetra_MultiVector *& b,
Epetra_MultiVector *& bt,
Epetra_MultiVector *&xexact, bool StaticProfile, bool MakeLocalOnly) {
Epetra_Time timer(comm);
// Determine my global IDs
int * myGlobalElements;
GenerateMyGlobalElements(numNodesX, numNodesY, numProcsX, numProcsY, comm.MyPID(), myGlobalElements);
int numMyEquations = numNodesX*numNodesY;
map = new Epetra_Map(-1, numMyEquations, myGlobalElements, 0, comm); // Create map with 2D block partitioning.
delete [] myGlobalElements;
int numGlobalEquations = map->NumGlobalElements();
int profile = 0; if (StaticProfile) profile = numPoints;
#ifdef EPETRA_HAVE_STATICPROFILE
if (MakeLocalOnly)
A = new Epetra_CrsMatrix(Copy, *map, *map, profile, StaticProfile); // Construct matrix with rowmap=colmap
else
A = new Epetra_CrsMatrix(Copy, *map, profile, StaticProfile); // Construct matrix
#else
if (MakeLocalOnly)
A = new Epetra_CrsMatrix(Copy, *map, *map, profile); // Construct matrix with rowmap=colmap
else
A = new Epetra_CrsMatrix(Copy, *map, profile); // Construct matrix
#endif
int * indices = new int[numPoints];
double * values = new double[numPoints];
double dnumPoints = (double) numPoints;
int nx = numNodesX*numProcsX;
for (int i=0; i<numMyEquations; i++) {
int rowID = map->GID(i);
int numIndices = 0;
for (int j=0; j<numPoints; j++) {
int colID = rowID + xoff[j] + nx*yoff[j]; // Compute column ID based on stencil offsets
if (colID>-1 && colID<numGlobalEquations) {
indices[numIndices] = colID;
double value = - ((double) rand())/ ((double) RAND_MAX);
if (colID==rowID)
values[numIndices++] = dnumPoints - value; // Make diagonal dominant
else
values[numIndices++] = value;
}
}
//cout << "Building row " << rowID << endl;
A->InsertGlobalValues(rowID, numIndices, values, indices);
}
delete [] indices;
delete [] values;
double insertTime = timer.ElapsedTime();
timer.ResetStartTime();
A->FillComplete(false);
double fillCompleteTime = timer.ElapsedTime();
if (verbose)
cout << "Time to insert matrix values = " << insertTime << endl
<< "Time to complete fill = " << fillCompleteTime << endl;
if (summary) {
if (comm.NumProc()==1) cout << "InsertTime" << '\t';
cout << insertTime << endl;
if (comm.NumProc()==1) cout << "FillCompleteTime" << '\t';
cout << fillCompleteTime << endl;
}
if (nrhs<=1) {
b = new Epetra_Vector(*map);
bt = new Epetra_Vector(*map);
xexact = new Epetra_Vector(*map);
}
else {
b = new Epetra_MultiVector(*map, nrhs);
bt = new Epetra_MultiVector(*map, nrhs);
xexact = new Epetra_MultiVector(*map, nrhs);
}
xexact->Random(); // Fill xexact with random values
A->Multiply(false, *xexact, *b);
A->Multiply(true, *xexact, *bt);
return;
}
int main(int argc, char *argv[]) {
#ifdef EPETRA_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
int MyPID = Comm.MyPID();
bool verbose = true;
if (MyPID==0) verbose = true;
if (verbose)
cout << EpetraExt::EpetraExt_Version() << endl << endl;
cout << Comm << endl;
if(argc < 2 && verbose) {
cerr << "Usage: " << argv[0]
<< " HB_filename" << endl;
return(1);
}
// Uncomment the next three lines to debug in mpi mode
//int tmp;
//if (MyPID==0) cin >> tmp;
//Comm.Barrier();
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra(argv[1], Comm, readMap, readA, readx, readb, readxexact);
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
Epetra_Vector x(map);
Epetra_Vector b(map);
Epetra_Vector xexact(map);
Epetra_Time FillTimer(Comm);
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
double vectorRedistributeTime = FillTimer.ElapsedTime();
A.Export(*readA, exporter, Add);
Comm.Barrier();
double matrixRedistributeTime = FillTimer.ElapsedTime() - vectorRedistributeTime;
assert(A.FillComplete()==0);
Comm.Barrier();
double fillCompleteTime = FillTimer.ElapsedTime() - matrixRedistributeTime;
if (Comm.MyPID()==0) {
cout << "\n\n****************************************************" << endl;
cout << "\n Vector redistribute time (sec) = " << vectorRedistributeTime<< endl;
cout << " Matrix redistribute time (sec) = " << matrixRedistributeTime << endl;
cout << " Transform to Local time (sec) = " << fillCompleteTime << endl<< endl;
}
Epetra_Vector tmp1(*readMap);
Epetra_Vector tmp2(map);
readA->Multiply(false, *readxexact, tmp1);
A.Multiply(false, xexact, tmp2);
double residual;
tmp1.Norm2(&residual);
if (verbose) cout << "Norm of Ax from file = " << residual << endl;
tmp2.Norm2(&residual);
if (verbose) cout << "Norm of Ax after redistribution = " << residual << endl << endl << endl;
//cout << "A from file = " << *readA << endl << endl << endl;
//cout << "A after dist = " << A << endl << endl << endl;
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
Comm.Barrier();
EpetraExt::RowMatrixToMatrixMarketFile("test.mm", A, "test matrix", "This is a test matrix");
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return 0 ;
}
/* Find the DBBD form */
int shylu_symbolic_factor
(
Epetra_CrsMatrix *A, // i/p: A matrix
shylu_symbolic *ssym, // symbolic structure
shylu_data *data, // numeric structure, TODO: Required ?
shylu_config *config // i/p: library configuration
)
{
#ifdef TIMING_OUTPUT
Teuchos::Time symtime("symbolic time");
symtime.start();
#endif
int myPID = A->Comm().MyPID();
int n = A->NumGlobalRows();
int Dnr;
int Snr;
int *DRowElems;
int *SRowElems;
int sym = config->sym;
checkMaps(A);
// Get column map
Epetra_Map AColMap = A->ColMap();
int ncols = AColMap.NumMyElements();
int *cols = AColMap.MyGlobalElements();
// Get row map
Epetra_Map ARowMap = A->RowMap();
int nrows = ARowMap.NumMyElements();
int *rows = ARowMap.MyGlobalElements();
// Find all columns in this proc
int *gvals = new int[n]; // vector of size n, not ncols !
// gvals[local cols] = 1, gvals[shared cols] > 1.
int SNumGlobalCols;
findLocalColumns(A, gvals, SNumGlobalCols);
// See if you can shrink the separator by assigning more rows/columns to
// the block diagonals
// TODO: This is because of a bug in coloring remove the if once that is
// fixed
//if (config->schurApproxMethod == 2)
if (config->sep_type == 2)
findNarrowSeparator(A, gvals);
// 3. Assemble diagonal block and the border in convenient form [
/* In each processor, we have (in a permuted form)
* | D_i C_i |
* | R_i S_i |
* D_i - diagonal block, C_i - Column Separator, R_i - Row separator
* S_i - A22 block corresponding to Schur complement part of A
* Assemble all four blocks in local matrices. */
ostringstream ssmsg1;
ssmsg1 << "PID =" << myPID << " ";
string msg = ssmsg1.str();
ssmsg1.clear(); ssmsg1.str("");
// Find #cols in each block
int Dnc = 0; // #cols in diagonal block
int Snc = 0; // #cols in the col. separator
/* Looping on cols will work only for wide separator
* as for narrow sep there will be some sep cols with gvals[col] ==1
* */
/*for (int i = 0; i < ncols ; i++)
{
if (gvals[cols[i]] == 1)
Dnc++;
else
Snc++;
}
// Find #rows in each block
Dnr = Dnc; // #rows in square diagonal block
Snr = nrows - Dnr; // #rows in the row separator*/
// Find #rows in each block
Dnr = 0;
Snr = 0;
for (int i = 0; i < nrows ; i++)
{
if (gvals[rows[i]] == 1)
Dnr++;
else
Snr++;
}
Dnc = Dnr;
// TODO: Snc is no longer useful, should remove it
for (int i = 0; i < ncols ; i++)
{
if (gvals[cols[i]] != 1)
Snc++;
}
assert(Snc >= 0);
// TODO : The above assignment may not be correct in the unsymetric case
////config->dm.print(2, msg + " Mycols=");
cout << msg << " Mycols="<< ncols << "Myrows ="<< nrows << endl;
cout << msg << " #rows and #cols in diagonal blk ="<< Dnr << endl;
cout << msg << " #columns in S ="<< Snc << endl;
cout << msg << " #rows in S ="<< Snr << endl;
ostringstream pidstr;
pidstr << myPID ;
// Create a row map for the D and S blocks [
DRowElems = new int[Dnr];
SRowElems = new int[Snr];
int gid;
// Assemble row ids in two arrays (for D and R blocks)
if (sym)
{
findBlockElems(A, nrows, rows, gvals, Dnr, DRowElems, Snr, SRowElems,
"D"+pidstr.str()+"Rows", "S"+pidstr.str()+"Rows", false) ;
}
else
{
// SRowElems are not known until factorization, TODO
assert(0 == 1);
}
data->Dnr = Dnr;
data->Snr = Snr;
data->Dnc = Dnc;
data->DRowElems = DRowElems;
data->SRowElems = SRowElems;
// Create a column map for the D and S blocks [
int *DColElems = new int[Dnc]; // Elems in column map of D
int *SColElems = new int[Snc]; // Elems in column map of C TODO: Unused
// Assemble column ids in two arrays (for D and C blocks)
findBlockElems(A, ncols, cols, gvals, Dnc, DColElems, Snc, SColElems,
"D"+pidstr.str()+"Cols", "S"+pidstr.str()+"Cols", true) ;
data->DColElems = DColElems;
data->gvals = gvals;
for (int i = 0; i < Snr; i++)
{
// Epetra guarentees columns corresponding to local rows will be first
// in the column map.
assert(SRowElems[i] == SColElems[i]);
}
// ]
/*--Create the Epetra Matrices with the maps (does not insert values) --- */
create_matrices(A, ssym, data, config);
/*--Extract the Epetra Matrices and call fillComplete --- */
extract_matrices(A, ssym, data, config, true);
delete[] SColElems;
Amesos Factory;
const char* SolverType = config->diagonalBlockSolver.c_str();
bool IsAvailable = Factory.Query(SolverType);
assert(IsAvailable == true);
Teuchos::RCP<Epetra_LinearProblem> LP = Teuchos::RCP<Epetra_LinearProblem>
(new Epetra_LinearProblem());
LP->SetOperator((ssym->D).getRawPtr());
//LP->SetOperator((ssym->DT).getRawPtr()); // for transpose
// Create temp vectors
ssym->Dlhs = Teuchos::RCP<Epetra_MultiVector>
(new Epetra_MultiVector(ssym->D->RowMap(), 16));
ssym->Drhs = Teuchos::RCP<Epetra_MultiVector>
(new Epetra_MultiVector(ssym->D->RowMap(), 16));
ssym->Gvec = Teuchos::RCP<Epetra_MultiVector>
(new Epetra_MultiVector(ssym->G->RowMap(), 16));
LP->SetRHS(ssym->Drhs.getRawPtr());
LP->SetLHS(ssym->Dlhs.getRawPtr());
ssym->ReIdx_LP = Teuchos::RCP<
EpetraExt::ViewTransform<Epetra_LinearProblem> >
(new EpetraExt::LinearProblem_Reindex2(0));
ssym->LP = Teuchos::RCP<Epetra_LinearProblem>(&((*(ssym->ReIdx_LP))(*LP)),
false);
Teuchos::RCP<Amesos_BaseSolver> Solver = Teuchos::RCP<Amesos_BaseSolver>
(Factory.Create(SolverType, *(ssym->LP)));
//config->dm.print(5, "Created the diagonal solver");
#ifdef TIMING_OUTPUT
Teuchos::Time ftime("setup time");
ftime.start();
#endif
//Solver->SetUseTranspose(true); // for transpose
Teuchos::ParameterList aList;
aList.set("TrustMe", true);
Solver->SetParameters(aList);
Solver->SymbolicFactorization();
//config->dm.print(3, "Symbolic Factorization done");
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Symbolic Factorization Time" << ftime.totalElapsedTime() << endl;
ftime.reset();
#endif
ssym->OrigLP = LP;
//ssym->LP = LP;
ssym->Solver = Solver;
if (config->schurApproxMethod == 1)
{
Teuchos::ParameterList pList;
Teuchos::RCP<Isorropia::Epetra::Prober> prober =
Teuchos::RCP<Isorropia::Epetra::Prober> (new
Isorropia::Epetra::Prober((ssym->Sg).getRawPtr(),
pList, false));
//config->dm.print(3, "Doing Coloring");
#ifdef TIMING_OUTPUT
ftime.start();
#endif
prober->color();
#ifdef TIMING_OUTPUT
ftime.stop();
cout << "Time to color" << ftime.totalElapsedTime() << endl;
ftime.reset();
ftime.start();
#endif
ssym->prober = prober;
}
#ifdef TIMING_OUTPUT
symtime.stop();
cout << "Symbolic Time" << symtime.totalElapsedTime() << endl;
symtime.reset();
#endif
}
int CreateTridi(Epetra_CrsMatrix& A)
{
Epetra_Map Map = A.RowMap();
int NumMyElements = Map.NumMyElements();
int NumGlobalElements = Map.NumGlobalElements();
int * MyGlobalElements = new int[NumMyElements];
Map.MyGlobalElements(MyGlobalElements);
// Add rows one-at-a-time
// Need some vectors to help
// Off diagonal Values will always be -1
double *Values = new double[3];
int *Indices = new int[3];
int NumEntries;
for (int i=0; i<NumMyElements; i++)
{
if (MyGlobalElements[i]==0)
{
Indices[0] = 0;
Indices[1] = 1;
Values[0] = 2.0;
Values[1] = -1.0;
NumEntries = 2;
}
else if (MyGlobalElements[i] == NumGlobalElements-1)
{
Indices[0] = NumGlobalElements-1;
Indices[1] = NumGlobalElements-2;
Values[0] = 2.0;
Values[1] = -1.0;
NumEntries = 2;
}
else
{
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i];
Indices[2] = MyGlobalElements[i]+1;
Values[0] = -1.0;
Values[1] = 2.0;
Values[2] = -1.0;
NumEntries = 3;
}
assert(A.InsertGlobalValues(MyGlobalElements[i], NumEntries, Values, Indices)==0);
// Put in the diagonal entry
// assert(A.InsertGlobalValues(MyGlobalElements[i], 1, &two, &MyGlobalElements[i])==0);
}
// Finish up
assert(A.FillComplete()==0);
delete[] MyGlobalElements;
delete[] Values;
delete[] Indices;
return 0;
}
int Drumm3(const Epetra_Map& map, bool verbose)
{
const Epetra_Comm & Comm = map.Comm();
/* get number of processors and the name of this processor */
int Numprocs = Comm.NumProc();
int MyPID = Comm.MyPID();
if (Numprocs != 2) return(0);
int NumGlobalRows = 4;
int IndexBase = 0;
Epetra_Map Map(NumGlobalRows, IndexBase, Comm);
// Construct FECrsMatrix
int NumEntriesPerRow = 3;
Epetra_FECrsMatrix A(Copy, Map, NumEntriesPerRow);
double ElementArea = 0.5;
int NumCols = 3;
int* Indices = new int[NumCols];
if(MyPID==0) // indices corresponding to element 0 on processor 0
{
Indices[0] = 0;
Indices[1] = 1;
Indices[2] = 3;
}
else if(MyPID==1) // indices corresponding to element 1 on processor 1
{
Indices[0] = 1;
Indices[1] = 2;
Indices[2] = 3;
}
double* Values = new double[NumCols*NumCols];
// removal term
Values[0] = 2*ElementArea/12.;
Values[1] = 1*ElementArea/12.;
Values[2] = 1*ElementArea/12.;
Values[3] = 1*ElementArea/12.;
Values[4] = 2*ElementArea/12.;
Values[5] = 1*ElementArea/12.;
Values[6] = 1*ElementArea/12.;
Values[7] = 1*ElementArea/12.;
Values[8] = 2*ElementArea/12.;
A.InsertGlobalValues(NumCols, Indices,
Values,
Epetra_FECrsMatrix::ROW_MAJOR);
A.GlobalAssemble();
A.GlobalAssemble();
// A.Print(cout);
// Create vectors for CG algorithm
Epetra_FEVector* bptr = new Epetra_FEVector(A.RowMap(), 1);
Epetra_FEVector* x0ptr = new Epetra_FEVector(A.RowMap(), 1);
Epetra_FEVector& b = *bptr;
Epetra_FEVector& x0 = *x0ptr;
// source terms
NumCols = 2;
if(MyPID==0) // indices corresponding to element 0 on processor 0
{
Indices[0] = 0;
Indices[1] = 3;
Values[0] = 1./2.;
Values[1] = 1./2.;
}
else
{
Indices[0] = 1;
Indices[1] = 2;
Values[0] = 0;
Values[1] = 0;
}
b.SumIntoGlobalValues(NumCols, Indices, Values);
b.GlobalAssemble();
if (verbose&&MyPID==0) cout << "b:" << endl;
if (verbose) {
b.Print(cout);
}
x0 = b;
if (verbose&&MyPID==0) {
cout << "x:"<<endl;
}
if (verbose) {
x0.Print(cout);
}
delete [] Values;
delete [] Indices;
delete bptr;
delete x0ptr;
return(0);
}
int checkmap(Epetra_Map & Map, int NumGlobalElements, int NumMyElements,
int *MyGlobalElements, int IndexBase, Epetra_Comm& Comm,
bool DistributedGlobal)
{
int i, ierr=0, forierr = 0;
EPETRA_TEST_ERR(!Map.ConstantElementSize(),ierr);
EPETRA_TEST_ERR(DistributedGlobal!=Map.DistributedGlobal(),ierr);
EPETRA_TEST_ERR(Map.ElementSize()!=1,ierr);
int *MyElementSizeList = new int[NumMyElements];
EPETRA_TEST_ERR(Map.ElementSizeList(MyElementSizeList)!=0,ierr);
forierr = 0;
for (i=0; i<NumMyElements; i++) forierr += MyElementSizeList[i]!=1;
EPETRA_TEST_ERR(forierr,ierr);
delete [] MyElementSizeList;
const Epetra_Comm & Comm1 = Map.Comm();
EPETRA_TEST_ERR(Comm1.NumProc()!=Comm.NumProc(),ierr);
EPETRA_TEST_ERR(Comm1.MyPID()!=Comm.MyPID(),ierr);
EPETRA_TEST_ERR(Map.IndexBase()!=IndexBase,ierr);
EPETRA_TEST_ERR(!Map.LinearMap() && MyGlobalElements==0,ierr);
EPETRA_TEST_ERR(Map.LinearMap() && MyGlobalElements!=0,ierr);
EPETRA_TEST_ERR(Map.MaxAllGID()!=NumGlobalElements-1+IndexBase,ierr);
EPETRA_TEST_ERR(Map.MaxElementSize()!=1,ierr);
int MaxLID = Map.MaxLID();
EPETRA_TEST_ERR(MaxLID!=NumMyElements-1,ierr);
int MaxMyGID = (Comm.MyPID()+1)*NumMyElements-1+IndexBase;
if (Comm.MyPID()>2) MaxMyGID+=3;
if (!DistributedGlobal) MaxMyGID = NumMyElements-1+IndexBase;
EPETRA_TEST_ERR(Map.MaxMyGID()!=MaxMyGID,ierr);
EPETRA_TEST_ERR(Map.MinAllGID()!=IndexBase,ierr);
EPETRA_TEST_ERR(Map.MinElementSize()!=1,ierr);
EPETRA_TEST_ERR(Map.MinLID()!=0,ierr);
int MinMyGID = Comm.MyPID()*NumMyElements+IndexBase;
if (Comm.MyPID()>2) MinMyGID+=3;
if (!DistributedGlobal) MinMyGID = 0;
EPETRA_TEST_ERR(Map.MinMyGID()!=MinMyGID,ierr);
int * MyGlobalElements1 = new int[NumMyElements];
EPETRA_TEST_ERR(Map.MyGlobalElements(MyGlobalElements1)!=0,ierr);
forierr = 0;
if (MyGlobalElements==0)
{
for (i=0; i<NumMyElements; i++)
forierr += MyGlobalElements1[i]!=MinMyGID+i;
EPETRA_TEST_ERR(forierr,ierr);
}
else {
for (i=0; i<NumMyElements; i++)
forierr += MyGlobalElements[i]!=MyGlobalElements1[i];
EPETRA_TEST_ERR(forierr,ierr);
}
EPETRA_TEST_ERR(Map.NumGlobalElements()!=NumGlobalElements,ierr);
EPETRA_TEST_ERR(Map.NumGlobalPoints()!=NumGlobalElements,ierr);
EPETRA_TEST_ERR(Map.NumMyElements()!=NumMyElements,ierr);
EPETRA_TEST_ERR(Map.NumMyPoints()!=NumMyElements,ierr);
int MaxMyGID2 = Map.GID(Map.LID(MaxMyGID));
EPETRA_TEST_ERR(MaxMyGID2 != MaxMyGID,ierr);
int MaxLID2 = Map.LID(Map.GID(MaxLID));
EPETRA_TEST_ERR(MaxLID2 != MaxLID,ierr);
EPETRA_TEST_ERR(Map.GID(MaxLID+1) != IndexBase-1,ierr);// MaxLID+1 doesn't exist
EPETRA_TEST_ERR(Map.LID(MaxMyGID+1) != -1,ierr);// MaxMyGID+1 doesn't exist or is on a different processor
EPETRA_TEST_ERR(!Map.MyGID(MaxMyGID),ierr);
EPETRA_TEST_ERR(Map.MyGID(MaxMyGID+1),ierr);
EPETRA_TEST_ERR(!Map.MyLID(MaxLID),ierr);
EPETRA_TEST_ERR(Map.MyLID(MaxLID+1),ierr);
EPETRA_TEST_ERR(!Map.MyGID(Map.GID(MaxLID)),ierr);
EPETRA_TEST_ERR(Map.MyGID(Map.GID(MaxLID+1)),ierr);
EPETRA_TEST_ERR(!Map.MyLID(Map.LID(MaxMyGID)),ierr);
EPETRA_TEST_ERR(Map.MyLID(Map.LID(MaxMyGID+1)),ierr);
// Check RemoteIDList function
// Get some GIDs off of each processor to test
int TotalNumEle, NumElePerProc, NumProc = Comm.NumProc();
int MinNumEleOnProc;
int NumMyEle=Map.NumMyElements();
Comm.MinAll(&NumMyEle,&MinNumEleOnProc,1);
if (MinNumEleOnProc > 5) NumElePerProc = 6;
else NumElePerProc = MinNumEleOnProc;
if (NumElePerProc > 0) {
TotalNumEle = NumElePerProc*NumProc;
int * MyGIDlist = new int[NumElePerProc];
int * GIDlist = new int[TotalNumEle];
int * PIDlist = new int[TotalNumEle];
int * LIDlist = new int[TotalNumEle];
for (i=0; i<NumElePerProc; i++)
MyGIDlist[i] = MyGlobalElements1[i];
Comm.GatherAll(MyGIDlist,GIDlist,NumElePerProc);// Get a few values from each proc
Map.RemoteIDList(TotalNumEle, GIDlist, PIDlist, LIDlist);
int MyPID= Comm.MyPID();
forierr = 0;
for (i=0; i<TotalNumEle; i++) {
if (Map.MyGID(GIDlist[i])) {
forierr += PIDlist[i] != MyPID;
forierr += !Map.MyLID(Map.LID(GIDlist[i])) || Map.LID(GIDlist[i]) != LIDlist[i] || Map.GID(LIDlist[i]) != GIDlist[i];
}
else {
forierr += PIDlist[i] == MyPID; // If MyGID comes back false, the PID listed should be that of another proc
}
}
EPETRA_TEST_ERR(forierr,ierr);
delete [] MyGIDlist;
delete [] GIDlist;
delete [] PIDlist;
delete [] LIDlist;
}
delete [] MyGlobalElements1;
// Check RemoteIDList function (assumes all maps are linear, even if not stored that way)
if (Map.LinearMap()) {
int * GIDList = new int[3];
int * PIDList = new int[3];
int * LIDList = new int[3];
int MyPID = Map.Comm().MyPID();
int NumIDs = 0;
//GIDList[NumIDs++] = Map.MaxAllGID()+1; // Should return -1 for both PID and LID
if (Map.MinMyGID()-1>=Map.MinAllGID()) GIDList[NumIDs++] = Map.MinMyGID()-1;
if (Map.MaxMyGID()+1<=Map.MaxAllGID()) GIDList[NumIDs++] = Map.MaxMyGID()+1;
Map.RemoteIDList(NumIDs, GIDList, PIDList, LIDList);
NumIDs = 0;
//EPETRA_TEST_ERR(!(PIDList[NumIDs]==-1),ierr);
//EPETRA_TEST_ERR(!(LIDList[NumIDs++]==-1),ierr);
if (Map.MinMyGID()-1>=Map.MinAllGID()) EPETRA_TEST_ERR(!(PIDList[NumIDs++]==MyPID-1),ierr);
if (Map.MaxMyGID()+1<=Map.MaxAllGID()) EPETRA_TEST_ERR(!(PIDList[NumIDs]==MyPID+1),ierr);
if (Map.MaxMyGID()+1<=Map.MaxAllGID()) EPETRA_TEST_ERR(!(LIDList[NumIDs++]==0),ierr);
delete [] GIDList;
delete [] PIDList;
delete [] LIDList;
}
return (ierr);
}
int Drumm1(const Epetra_Map& map, bool verbose)
{
(void)verbose;
//Simple 2-element problem (element as in "finite-element") from
//Clif Drumm. Two triangular elements, one per processor, as shown
//here:
//
// *----*
// 3|\ 2|
// | \ |
// | 0\1|
// | \|
// *----*
// 0 1
//
//Element 0 on processor 0, element 1 on processor 1.
//Processor 0 will own nodes 0,1 and processor 1 will own nodes 2,3.
//Each processor will pass a 3x3 element-matrix to Epetra_FECrsMatrix.
//After GlobalAssemble(), the matrix should be as follows:
//
// row 0: 2 1 0 1
//proc 0 row 1: 1 4 1 2
//----------------------------------
// row 2: 0 1 2 1
//proc 1 row 3: 1 2 1 4
//
int numProcs = map.Comm().NumProc();
int localProc = map.Comm().MyPID();
if (numProcs != 2) return(0);
//so first we'll set up a epetra_test::matrix_data object with
//contents that match the above-described matrix. (but the
//matrix_data object will have all 4 rows on each processor)
int i;
int rowlengths[4];
rowlengths[0] = 3;
rowlengths[1] = 4;
rowlengths[2] = 3;
rowlengths[3] = 4;
epetra_test::matrix_data matdata(4, rowlengths);
for(i=0; i<4; ++i) {
for(int j=0; j<matdata.rowlengths()[i]; ++j) {
matdata.colindices()[i][j] = j;
}
}
matdata.colindices()[0][2] = 3;
matdata.colindices()[2][0] = 1;
matdata.colindices()[2][1] = 2;
matdata.colindices()[2][2] = 3;
double** coefs = matdata.coefs();
coefs[0][0] = 2.0; coefs[0][1] = 1.0; coefs[0][2] = 1.0;
coefs[1][0] = 1.0; coefs[1][1] = 4.0; coefs[1][2] = 1.0; coefs[1][3] = 2.0;
coefs[2][0] = 1.0; coefs[2][1] = 2.0; coefs[2][2] = 1.0;
coefs[3][0] = 1.0; coefs[3][1] = 2.0; coefs[3][2] = 1.0; coefs[3][3] = 4.0;
//now we'll load a Epetra_FECrsMatrix with data that matches the
//above-described finite-element problem.
int indexBase = 0, ierr = 0;
int myNodes[4];
double values[9];
values[0] = 2.0;
values[1] = 1.0;
values[2] = 1.0;
values[3] = 1.0;
values[4] = 2.0;
values[5] = 1.0;
values[6] = 1.0;
values[7] = 1.0;
values[8] = 2.0;
int numMyNodes = 2;
if (localProc == 0) {
myNodes[0] = 0;
myNodes[1] = 1;
}
else {
myNodes[0] = 2;
myNodes[1] = 3;
}
Epetra_Map Map(-1, numMyNodes, myNodes, indexBase, map.Comm());
numMyNodes = 3;
if (localProc == 0) {
myNodes[0] = 0;
myNodes[1] = 1;
myNodes[2] = 3;
}
else {
myNodes[0] = 1;
myNodes[1] = 2;
myNodes[2] = 3;
}
int rowLengths = 3;
Epetra_FECrsMatrix A(Copy, Map, rowLengths);
EPETRA_TEST_ERR( A.InsertGlobalValues(numMyNodes, myNodes,
numMyNodes, myNodes, values,
Epetra_FECrsMatrix::ROW_MAJOR),ierr);
EPETRA_TEST_ERR( A.GlobalAssemble(), ierr );
EPETRA_TEST_ERR( A.GlobalAssemble(), ierr );
//now the test is to check whether the FECrsMatrix data matches the
//epetra_test::matrix_data object...
bool the_same = matdata.compare_local_data(A);
if (!the_same) {
return(-1);
}
return(0);
}
int main(int argc, char *argv[])
{
#ifdef HAVE_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm( MPI_COMM_WORLD );
#else
Epetra_SerialComm Comm;
#endif
// My MPI process rank.
const int MyPID = Comm.MyPID();
// "/Users/sakashitatatsuya/Downloads/barista_trunk_slepc/sample/hamiltonian_matrix.ip"
std::ifstream ifs(argv[1]);
alps::Parameters params(ifs);
Teuchos::oblackholestream blackHole;
std::ostream& out = (MyPID == 0) ? std::cout : blackHole;
barista::Hamiltonian<> hamiltonian(params);
matrix_type matrix(hamiltonian.dimension(), hamiltonian.dimension());
hamiltonian.fill<double>(matrix);
int m,n;
int N;
m = n = N = hamiltonian.dimension();
//std::cout << matrix << std::endl;
std::ofstream ofs;
if (MyPID==0) {
ofs.open("anasazi_time.txt");
if (!ofs) {
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return -1;
}
}
//Teuchos::ParameterList GaleriList;
using Teuchos::RCP;
using Teuchos::rcp;
typedef Teuchos::ScalarTraits<double> STS;
const double one = STS::one();
const double zero = STS::zero();
// The problem is defined on a 2D grid, global size is nx * nx.
//int nx = N;
//GaleriList.set("n", nx * nx);
//GaleriList.set("nx", nx);
//GaleriList.set("ny", nx);
//Teuchos::RCP<Epetra_Map> Map = Teuchos::rcp( Galeri::CreateMap("Linear", Comm, GaleriList) );
//Teuchos::RCP<Epetra_RowMatrix> A = Teuchos::rcp( Galeri::CreateCrsMatrix("Laplace2D", &*Map, GaleriList) );
// Construct a Map that puts approximately the same number of rows
// of the matrix A on each processor.
Epetra_Map RowMap (N, 0, Comm);
Epetra_Map ColMap (N, 0, Comm);
// Get update list and number of local equations from newly created Map.
const int NumMyRowElements = RowMap.NumMyElements ();
std::vector<int> MyGlobalRowElements (NumMyRowElements);
RowMap.MyGlobalElements (&MyGlobalRowElements[0]);
// Create an Epetra_CrsMatrix using the given row map.
RCP<Epetra_CrsMatrix> A = rcp (new Epetra_CrsMatrix (Copy, RowMap, n));
// We use info to catch any errors that may have happened during // matrix assembly, and report them globally. We do this so that // the MPI processes won't call FillComplete() unless they all // successfully filled their parts of the matrix.
int info = 0;
try {
//
// Compute coefficients for the discrete integral operator.
//
std::vector<double> Values (n);
std::vector<int> Indices (n);
//const double inv_mp1 = one / (m+1);
//const double inv_np1 = one / (n+1);
int count;
//for (int i = 0; i < n; ++i) {
// Indices[i] = i;
//}
for (int i = 0; i < NumMyRowElements; ++i) {
count =0;
for (int j = 0; j < n; ++j) {
if (matrix(MyGlobalRowElements[i],j)!=0) {
Values[count] = matrix(MyGlobalRowElements[i],j);
Indices[count] = j;
count++;
}
}
info = A->InsertGlobalValues (MyGlobalRowElements[i], count,
&Values[0], &Indices[0]);
// Make sure that the insertion succeeded. Teuchos'
// TEST_FOR_EXCEPTION macro gives a nice error message if the
// thrown exception isn't caught. We'll report this on the
// offending MPI process.
/*
TEST_FOR_EXCEPTION( info != 0, std::runtime_error, "Failed to insert n="
<< n << " global value" << (n != 1 ? "s" : "")
<< " in row " << MyGlobalRowElements[i]
<< " of the matrix." );
*/
} // for i = 0...
// Call FillComplete on the matrix. Since the matrix isn't square,
// we have to give FillComplete the domain and range maps, which in
// this case are the column resp. row maps.
info = A->FillComplete (ColMap, RowMap);
/*
TEST_FOR_EXCEPTION( info != 0, std::runtime_error,
"FillComplete failed with INFO = " << info << ".");
*/
info = A->OptimizeStorage();
/*
TEST_FOR_EXCEPTION( info != 0, std::runtime_error,
"OptimizeStorage failed with INFO = " << info << ".");
*/
} catch (std::runtime_error& e) {
// If multiple MPI processes are reporting errors, sometimes
// forming the error message as a string and then writing it to
// the output stream prevents messages from different processes
// from being interleaved.
std::ostringstream os;
os << "*** Error on MPI process " << MyPID << ": " << e.what();
cerr << os.str() << endl;
if (info == 0)
info = -1; // All procs will share info later on.
}
// Variables used for the Block Davidson Method
const int nev = 5;
const int blockSize = 5;
const int numBlocks = 8;
const int maxRestarts = 500;
const double tol = 1.0e-8;
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef Anasazi::MultiVecTraits<double, Epetra_MultiVector> MVT;
// Create an Epetra_MultiVector for an initial vector to start the solver.
// Note: This needs to have the same number of columns as the blocksize.
//
//Teuchos::RCP<Epetra_MultiVector> ivec = Teuchos::rcp( new Epetra_MultiVector(*Map, blockSize) );
Teuchos::RCP<Epetra_MultiVector> ivec = Teuchos::rcp( new Epetra_MultiVector(ColMap, blockSize) );
ivec->Random();
// Create the eigenproblem.
Teuchos::RCP<Anasazi::BasicEigenproblem<double, MV, OP> > problem =
Teuchos::rcp( new Anasazi::BasicEigenproblem<double, MV, OP>(A, ivec) );
// Inform the eigenproblem that the operator A is symmetric
problem->setHermitian(true);
// Set the number of eigenvalues requested
problem->setNEV( nev );
// Inform the eigenproblem that you are finishing passing it information
bool boolret = problem->setProblem();
if (boolret != true) {
std::cout<<"Anasazi::BasicEigenproblem::setProblem() returned an error." << std::endl;
#ifdef HAVE_MPI
MPI_Finalize();
#endif
return -1;
}
// Create parameter list to pass into the solver manager
Teuchos::ParameterList anasaziPL;
anasaziPL.set( "Which", "LM" );
anasaziPL.set( "Block Size", blockSize );
anasaziPL.set( "Maximum Iterations", 500 );
anasaziPL.set( "Convergence Tolerance", tol );
anasaziPL.set( "Verbosity", Anasazi::Errors+Anasazi::Warnings+Anasazi::TimingDetails+Anasazi::FinalSummary );
// Create the solver manager
Anasazi::LOBPCGSolMgr<double, MV, OP> anasaziSolver(problem, anasaziPL);
// Solve the problem
double start, end;
MPI_Barrier(MPI_COMM_WORLD);
start = MPI_Wtime();
Anasazi::ReturnType returnCode = anasaziSolver.solve();
MPI_Barrier(MPI_COMM_WORLD);
end = MPI_Wtime();
// Get the eigenvalues and eigenvectors from the eigenproblem
Anasazi::Eigensolution<double,MV> sol = problem->getSolution();
std::vector<Anasazi::Value<double> > evals = sol.Evals;
Teuchos::RCP<MV> evecs = sol.Evecs;
// Compute residuals.
std::vector<double> normR(sol.numVecs);
Teuchos::SerialDenseMatrix<int,double> T(sol.numVecs, sol.numVecs);
Epetra_MultiVector tempAevec( ColMap, sol.numVecs );
T.putScalar(0.0);
for (int i=0; i<sol.numVecs; i++) {
T(i,i) = evals[i].realpart;
}
A->Apply( *evecs, tempAevec );
MVT::MvTimesMatAddMv( -1.0, *evecs, T, 1.0, tempAevec );
MVT::MvNorm( tempAevec, normR );
if (MyPID == 0) {
// Print the results
std::cout<<"Solver manager returned " << (returnCode == Anasazi::Converged ? "converged." : "unconverged.") << std::endl;
std::cout<<std::endl;
std::cout<<"------------------------------------------------------"<<std::endl;
std::cout<<std::setw(16)<<"Eigenvalue"
<<std::setw(18)<<"Direct Residual"
<<std::endl;
std::cout<<"------------------------------------------------------"<<std::endl;
for (int i=0; i<sol.numVecs; i++) {
std::cout<<std::setw(16)<<evals[i].realpart
<<std::setw(18)<<normR[i]/evals[i].realpart
<<std::endl;
}
std::cout<<"------------------------------------------------------"<<std::endl;
}
// Print out the map and matrices
//ColMap.Print (out);
//A->Print (cout);
//RowMap.Print (cout);
double time;
int iter;
if (MyPID==0) {
iter = anasaziSolver.getNumIters();
Teuchos::Array<Teuchos::RCP<Teuchos::Time> > timer = anasaziSolver.getTimers();
Teuchos::RCP<Teuchos::Time> _timerSolve = timer[0];
cout << "timerSolve=" << _timerSolve << endl;
time = end - start;
cout << "time=" << time << endl;
ofs << "time=" << time << endl;
cout << "iter=" << iter << endl;
ofs << "iter=" << iter << endl;
}
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return 0;
}
int main(int argc, char *argv[]) {
#ifdef EPETRA_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
cout << Comm << endl;
int MyPID = Comm.MyPID();
bool verbose = false;
bool verbose1 = true;
if (MyPID==0) verbose = true;
if(argc < 2 && verbose) {
cerr << "Usage: " << argv[0]
<< " HB_filename [level_fill [level_overlap [absolute_threshold [ relative_threshold]]]]" << endl
<< "where:" << endl
<< "HB_filename - filename and path of a Harwell-Boeing data set" << endl
<< "level_fill - The amount of fill to use for ILU(k) preconditioner (default 0)" << endl
<< "level_overlap - The amount of overlap used for overlapping Schwarz subdomains (default 0)" << endl
<< "absolute_threshold - The minimum value to place on the diagonal prior to factorization (default 0.0)" << endl
<< "relative_threshold - The relative amount to perturb the diagonal prior to factorization (default 1.0)" << endl << endl
<< "To specify a non-default value for one of these parameters, you must specify all" << endl
<< " preceding values but not any subsequent parameters. Example:" << endl
<< "ifpackHpcSerialMsr.exe mymatrix.hpc 1 - loads mymatrix.hpc, uses level fill of one, all other values are defaults" << endl
<< endl;
return(1);
}
// Uncomment the next three lines to debug in mpi mode
//int tmp;
//if (MyPID==0) cin >> tmp;
//Comm.Barrier();
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra(argv[1], Comm, readMap, readA, readx, readb, readxexact);
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
Epetra_Vector x(map);
Epetra_Vector b(map);
Epetra_Vector xexact(map);
Epetra_Time FillTimer(Comm);
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
double vectorRedistributeTime = FillTimer.ElapsedTime();
A.Export(*readA, exporter, Add);
Comm.Barrier();
double matrixRedistributeTime = FillTimer.ElapsedTime() - vectorRedistributeTime;
assert(A.FillComplete()==0);
Comm.Barrier();
double fillCompleteTime = FillTimer.ElapsedTime() - matrixRedistributeTime;
if (Comm.MyPID()==0) {
cout << "\n\n****************************************************" << endl;
cout << "\n Vector redistribute time (sec) = " << vectorRedistributeTime<< endl;
cout << " Matrix redistribute time (sec) = " << matrixRedistributeTime << endl;
cout << " Transform to Local time (sec) = " << fillCompleteTime << endl<< endl;
}
Epetra_Vector tmp1(*readMap);
Epetra_Vector tmp2(map);
readA->Multiply(false, *readxexact, tmp1);
A.Multiply(false, xexact, tmp2);
double residual;
tmp1.Norm2(&residual);
if (verbose) cout << "Norm of Ax from file = " << residual << endl;
tmp2.Norm2(&residual);
if (verbose) cout << "Norm of Ax after redistribution = " << residual << endl << endl << endl;
//cout << "A from file = " << *readA << endl << endl << endl;
//cout << "A after dist = " << A << endl << endl << endl;
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
Comm.Barrier();
bool smallProblem = false;
if (A.RowMap().NumGlobalElements()<100) smallProblem = true;
if (smallProblem)
cout << "Original Matrix = " << endl << A << endl;
x.PutScalar(0.0);
Epetra_LinearProblem FullProblem(&A, &x, &b);
double normb, norma;
b.NormInf(&normb);
norma = A.NormInf();
if (verbose)
cout << "Inf norm of Original Matrix = " << norma << endl
<< "Inf norm of Original RHS = " << normb << endl;
Epetra_Time ReductionTimer(Comm);
Epetra_CrsSingletonFilter SingletonFilter;
Comm.Barrier();
double reduceInitTime = ReductionTimer.ElapsedTime();
SingletonFilter.Analyze(&A);
Comm.Barrier();
double reduceAnalyzeTime = ReductionTimer.ElapsedTime() - reduceInitTime;
if (SingletonFilter.SingletonsDetected())
cout << "Singletons found" << endl;
else {
cout << "Singletons not found" << endl;
exit(1);
}
SingletonFilter.ConstructReducedProblem(&FullProblem);
Comm.Barrier();
double reduceConstructTime = ReductionTimer.ElapsedTime() - reduceInitTime;
double totalReduceTime = ReductionTimer.ElapsedTime();
if (verbose)
cout << "\n\n****************************************************" << endl
<< " Reduction init time (sec) = " << reduceInitTime<< endl
<< " Reduction Analyze time (sec) = " << reduceAnalyzeTime << endl
<< " Construct Reduced Problem time (sec) = " << reduceConstructTime << endl
<< " Reduction Total time (sec) = " << totalReduceTime << endl<< endl;
Statistics(SingletonFilter);
Epetra_LinearProblem * ReducedProblem = SingletonFilter.ReducedProblem();
Epetra_CrsMatrix * Ap = dynamic_cast<Epetra_CrsMatrix *>(ReducedProblem->GetMatrix());
Epetra_Vector * bp = (*ReducedProblem->GetRHS())(0);
Epetra_Vector * xp = (*ReducedProblem->GetLHS())(0);
if (smallProblem)
cout << " Reduced Matrix = " << endl << *Ap << endl
<< " LHS before sol = " << endl << *xp << endl
<< " RHS = " << endl << *bp << endl;
// Construct ILU preconditioner
double elapsed_time, total_flops, MFLOPs;
Epetra_Time timer(Comm);
int LevelFill = 0;
if (argc > 2) LevelFill = atoi(argv[2]);
if (verbose) cout << "Using Level Fill = " << LevelFill << endl;
int Overlap = 0;
if (argc > 3) Overlap = atoi(argv[3]);
if (verbose) cout << "Using Level Overlap = " << Overlap << endl;
double Athresh = 0.0;
if (argc > 4) Athresh = atof(argv[4]);
if (verbose) cout << "Using Absolute Threshold Value of = " << Athresh << endl;
double Rthresh = 1.0;
if (argc > 5) Rthresh = atof(argv[5]);
if (verbose) cout << "Using Relative Threshold Value of = " << Rthresh << endl;
Ifpack_IlukGraph * IlukGraph = 0;
Ifpack_CrsRiluk * ILUK = 0;
if (LevelFill>-1) {
elapsed_time = timer.ElapsedTime();
IlukGraph = new Ifpack_IlukGraph(Ap->Graph(), LevelFill, Overlap);
assert(IlukGraph->ConstructFilledGraph()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
if (verbose) cout << "Time to construct ILUK graph = " << elapsed_time << endl;
Epetra_Flops fact_counter;
elapsed_time = timer.ElapsedTime();
ILUK = new Ifpack_CrsRiluk(*IlukGraph);
ILUK->SetFlopCounter(fact_counter);
ILUK->SetAbsoluteThreshold(Athresh);
ILUK->SetRelativeThreshold(Rthresh);
//assert(ILUK->InitValues()==0);
int initerr = ILUK->InitValues(*Ap);
if (initerr!=0) {
cout << endl << Comm << endl << " InitValues error = " << initerr;
if (initerr==1) cout << " Zero diagonal found, warning error only";
cout << endl << endl;
}
assert(ILUK->Factor()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = ILUK->Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute preconditioner values = "
<< elapsed_time << endl
<< "MFLOPS for Factorization = " << MFLOPs << endl;
//cout << *ILUK << endl;
double Condest;
ILUK->Condest(false, Condest);
if (verbose) cout << "Condition number estimate for this preconditioner = " << Condest << endl;
}
int Maxiter = 100;
double Tolerance = 1.0E-8;
Epetra_Flops counter;
Ap->SetFlopCounter(counter);
xp->SetFlopCounter(*Ap);
bp->SetFlopCounter(*Ap);
if (ILUK!=0) ILUK->SetFlopCounter(*Ap);
elapsed_time = timer.ElapsedTime();
double normreducedb, normreduceda;
bp->NormInf(&normreducedb);
normreduceda = Ap->NormInf();
if (verbose)
cout << "Inf norm of Reduced Matrix = " << normreduceda << endl
<< "Inf norm of Reduced RHS = " << normreducedb << endl;
BiCGSTAB(*Ap, *xp, *bp, ILUK, Maxiter, Tolerance, &residual, verbose);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = counter.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute solution = "
<< elapsed_time << endl
<< "Number of operations in solve = " << total_flops << endl
<< "MFLOPS for Solve = " << MFLOPs<< endl << endl;
SingletonFilter.ComputeFullSolution();
if (smallProblem)
cout << " Reduced LHS after sol = " << endl << *xp << endl
<< " Full LHS after sol = " << endl << x << endl
<< " Full Exact LHS = " << endl << xexact << endl;
Epetra_Vector resid(x);
resid.Update(1.0, x, -1.0, xexact, 0.0); // resid = xcomp - xexact
resid.Norm2(&residual);
double normx, normxexact;
x.Norm2(&normx);
xexact.Norm2(&normxexact);
if (verbose)
cout << "2-norm of computed solution = " << normx << endl
<< "2-norm of exact solution = " << normxexact << endl
<< "2-norm of difference between computed and exact solution = " << residual << endl;
if (verbose1 && residual>1.0e-5) {
if (verbose)
cout << "Difference between computed and exact solution appears large..." << endl
<< "Computing norm of A times this difference. If this norm is small, then matrix is singular"
<< endl;
Epetra_Vector bdiff(b);
assert(A.Multiply(false, resid, bdiff)==0);
assert(bdiff.Norm2(&residual)==0);
if (verbose)
cout << "2-norm of A times difference between computed and exact solution = " << residual << endl;
}
if (verbose)
cout << "********************************************************" << endl
<< " Solving again with 2*Ax=2*b" << endl
<< "********************************************************" << endl;
A.Scale(1.0); // A = 2*A
b.Scale(1.0); // b = 2*b
x.PutScalar(0.0);
b.NormInf(&normb);
norma = A.NormInf();
if (verbose)
cout << "Inf norm of Original Matrix = " << norma << endl
<< "Inf norm of Original RHS = " << normb << endl;
double updateReducedProblemTime = ReductionTimer.ElapsedTime();
SingletonFilter.UpdateReducedProblem(&FullProblem);
Comm.Barrier();
updateReducedProblemTime = ReductionTimer.ElapsedTime() - updateReducedProblemTime;
if (verbose)
cout << "\n\n****************************************************" << endl
<< " Update Reduced Problem time (sec) = " << updateReducedProblemTime<< endl
<< "****************************************************" << endl;
Statistics(SingletonFilter);
if (LevelFill>-1) {
Epetra_Flops fact_counter;
elapsed_time = timer.ElapsedTime();
int initerr = ILUK->InitValues(*Ap);
if (initerr!=0) {
cout << endl << Comm << endl << " InitValues error = " << initerr;
if (initerr==1) cout << " Zero diagonal found, warning error only";
cout << endl << endl;
}
assert(ILUK->Factor()==0);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = ILUK->Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute preconditioner values = "
<< elapsed_time << endl
<< "MFLOPS for Factorization = " << MFLOPs << endl;
double Condest;
ILUK->Condest(false, Condest);
if (verbose) cout << "Condition number estimate for this preconditioner = " << Condest << endl;
}
bp->NormInf(&normreducedb);
normreduceda = Ap->NormInf();
if (verbose)
cout << "Inf norm of Reduced Matrix = " << normreduceda << endl
<< "Inf norm of Reduced RHS = " << normreducedb << endl;
BiCGSTAB(*Ap, *xp, *bp, ILUK, Maxiter, Tolerance, &residual, verbose);
elapsed_time = timer.ElapsedTime() - elapsed_time;
total_flops = counter.Flops();
MFLOPs = total_flops/elapsed_time/1000000.0;
if (verbose) cout << "Time to compute solution = "
<< elapsed_time << endl
<< "Number of operations in solve = " << total_flops << endl
<< "MFLOPS for Solve = " << MFLOPs<< endl << endl;
SingletonFilter.ComputeFullSolution();
if (smallProblem)
cout << " Reduced LHS after sol = " << endl << *xp << endl
<< " Full LHS after sol = " << endl << x << endl
<< " Full Exact LHS = " << endl << xexact << endl;
resid.Update(1.0, x, -1.0, xexact, 0.0); // resid = xcomp - xexact
resid.Norm2(&residual);
x.Norm2(&normx);
xexact.Norm2(&normxexact);
if (verbose)
cout << "2-norm of computed solution = " << normx << endl
<< "2-norm of exact solution = " << normxexact << endl
<< "2-norm of difference between computed and exact solution = " << residual << endl;
if (verbose1 && residual>1.0e-5) {
if (verbose)
cout << "Difference between computed and exact solution appears large..." << endl
<< "Computing norm of A times this difference. If this norm is small, then matrix is singular"
<< endl;
Epetra_Vector bdiff(b);
assert(A.Multiply(false, resid, bdiff)==0);
assert(bdiff.Norm2(&residual)==0);
if (verbose)
cout << "2-norm of A times difference between computed and exact solution = " << residual << endl;
}
if (ILUK!=0) delete ILUK;
if (IlukGraph!=0) delete IlukGraph;
#ifdef EPETRA_MPI
MPI_Finalize() ;
#endif
return 0 ;
}
int
main (int argc, char *argv[])
{
using Teuchos::ArrayRCP;
using Teuchos::ArrayView;
using Teuchos::Comm;
using Teuchos::CommandLineProcessor;
using Teuchos::FancyOStream;
using Teuchos::getFancyOStream;
using Teuchos::OSTab;
using Teuchos::ptr;
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcpFromRef;
using std::cout;
using std::endl;
bool success = true; // May be changed by tests
Teuchos::oblackholestream blackHole;
//Teuchos::GlobalMPISession (&argc, &argv, &blackHole);
MPI_Init (&argc, &argv);
//
// Construct communicators, and verify that we are on 4 processors.
//
// Construct a Teuchos Comm object.
RCP<const Comm<int> > teuchosComm = Teuchos::DefaultComm<int>::getComm();
const int numProcs = teuchosComm->getSize();
const int pid = teuchosComm->getRank();
RCP<FancyOStream> pOut =
getFancyOStream (rcpFromRef ((pid == 0) ? std::cout : blackHole));
FancyOStream& out = *pOut;
// Verify that we are on four processors (which manifests the bug).
if (teuchosComm->getSize() != 4) {
out << "This test must be run on four processors. Exiting ..." << endl;
return EXIT_FAILURE;
}
// We also need an Epetra Comm, so that we can compare Tpetra and
// Epetra results.
Epetra_MpiComm epetraComm (MPI_COMM_WORLD);
//
// Default values of command-line options.
//
bool verbose = false;
bool printEpetra = false;
bool printTpetra = false;
CommandLineProcessor cmdp (false,true);
//
// Set command-line options.
//
cmdp.setOption ("verbose", "quiet", &verbose, "Print verbose output.");
// Epetra and Tpetra output will ask the Maps and Import objects to
// print themselves in distributed, maximally verbose fashion. It's
// best to turn on either Epetra or Tpetra, but not both. Then you
// can compare their output side by side.
cmdp.setOption ("printEpetra", "dontPrintEpetra", &printEpetra,
"Print Epetra output (in verbose mode only).");
cmdp.setOption ("printTpetra", "dontPrintTpetra", &printTpetra,
"Print Tpetra output (in verbose mode only).");
// Parse command-line options.
if (cmdp.parse (argc,argv) != CommandLineProcessor::PARSE_SUCCESSFUL) {
out << "End Result: TEST FAILED" << endl;
MPI_Finalize ();
return EXIT_FAILURE;
}
if (verbose) {
out << "Running test on " << numProcs << " process"
<< (numProcs != 1 ? "es" : "") << "." << endl;
}
// The maps for this problem are derived from a 3D structured mesh.
// In this example, the dimensions are 4x4x2 and there are 2
// processors assigned to the first dimension and 2 processors
// assigned to the second dimension, with no parallel decomposition
// along the third dimension. The "owned" arrays represent the
// one-to-one map, with each array representing a 2x2x2 slice. If
// DIMENSIONS == 2, then only the first 4 values will be used,
// representing a 2x2(x1) slice.
int owned0[8] = { 0, 1, 4, 5,16,17,20,21};
int owned1[8] = { 2, 3, 6, 7,18,19,22,23};
int owned2[8] = { 8, 9,12,13,24,25,28,29};
int owned3[8] = {10,11,14,15,26,27,30,31};
// The "overlap" arrays represent the map with communication
// elements, with each array representing a 3x3x2 slice. If
// DIMENSIONS == 2, then only the first 9 values will be used,
// representing a 3x3(x1) slice.
int overlap0[18] = {0,1,2,4, 5, 6, 8, 9,10,16,17,18,20,21,22,24,25,26};
int overlap1[18] = {1,2,3,5, 6, 7, 9,10,11,17,18,19,21,22,23,25,26,27};
int overlap2[18] = {4,5,6,8, 9,10,12,13,14,20,21,22,24,25,26,28,29,30};
int overlap3[18] = {5,6,7,9,10,11,13,14,15,21,22,23,25,26,27,29,30,31};
// Construct the owned and overlap maps for both Epetra and Tpetra.
int* owned;
int* overlap;
if (pid == 0) {
owned = owned0;
overlap = overlap0;
}
else if (pid == 1) {
owned = owned1;
overlap = overlap1;
}
else if (pid == 2) {
owned = owned2;
overlap = overlap2;
}
else {
owned = owned3;
overlap = overlap3;
}
#if DIMENSIONS == 2
int ownedSize = 4;
int overlapSize = 9;
#elif DIMENSIONS == 3
int ownedSize = 8;
int overlapSize = 18;
#endif
// Create the two Epetra Maps. Source for the Import is the owned
// map; target for the Import is the overlap map.
Epetra_Map epetraOwnedMap ( -1, ownedSize, owned, 0, epetraComm);
Epetra_Map epetraOverlapMap (-1, overlapSize, overlap, 0, epetraComm);
if (verbose && printEpetra) {
// Have the Epetra_Map objects describe themselves.
//
// Epetra_BlockMap::Print() takes an std::ostream&, and expects
// all MPI processes to be able to write to it. (The method
// handles its own synchronization.)
out << "Epetra owned map:" << endl;
epetraOwnedMap.Print (std::cout);
out << "Epetra overlap map:" << endl;
epetraOverlapMap.Print (std::cout);
}
// Create the two Tpetra Maps. The "invalid" global element count
// input tells Tpetra::Map to compute the global number of elements
// itself.
const int invalid = Teuchos::OrdinalTraits<int>::invalid();
RCP<Tpetra::Map<int> > tpetraOwnedMap =
rcp (new Tpetra::Map<int> (invalid, ArrayView<int> (owned, ownedSize),
0, teuchosComm));
tpetraOwnedMap->setObjectLabel ("Owned Map");
RCP<Tpetra::Map<int> > tpetraOverlapMap =
rcp (new Tpetra::Map<int> (invalid, ArrayView<int> (overlap, overlapSize),
0, teuchosComm));
tpetraOverlapMap->setObjectLabel ("Overlap Map");
// In verbose mode, have the Tpetra::Map objects describe themselves.
if (verbose && printTpetra) {
Teuchos::EVerbosityLevel verb = Teuchos::VERB_EXTREME;
// Tpetra::Map::describe() takes a FancyOStream, but expects all
// MPI processes to be able to write to it. (The method handles
// its own synchronization.)
RCP<FancyOStream> globalOut = getFancyOStream (rcpFromRef (std::cout));
out << "Tpetra owned map:" << endl;
{
OSTab tab (globalOut);
tpetraOwnedMap->describe (*globalOut, verb);
}
out << "Tpetra overlap map:" << endl;
{
OSTab tab (globalOut);
tpetraOverlapMap->describe (*globalOut, verb);
}
}
// Use the owned and overlap maps to construct an importer for both
// Epetra and Tpetra.
Epetra_Import epetraImporter (epetraOverlapMap, epetraOwnedMap );
Tpetra::Import<int> tpetraImporter (tpetraOwnedMap , tpetraOverlapMap);
// In verbose mode, have the Epetra_Import object describe itself.
if (verbose && printEpetra) {
out << "Epetra importer:" << endl;
// The importer's Print() method takes an std::ostream& and plans
// to write to it on all MPI processes (handling synchronization
// itself).
epetraImporter.Print (std::cout);
out << endl;
}
// In verbose mode, have the Tpetra::Import object describe itself.
if (verbose && printTpetra) {
out << "Tpetra importer:" << endl;
// The importer doesn't implement Teuchos::Describable. It wants
// std::cout and plans to write to it on all MPI processes (with
// its own synchronization).
tpetraImporter.print (std::cout);
out << endl;
}
// Construct owned and overlap vectors for both Epetra and Tpetra.
Epetra_Vector epetraOwnedVector (epetraOwnedMap );
Epetra_Vector epetraOverlapVector (epetraOverlapMap);
Tpetra::Vector<double,int> tpetraOwnedVector (tpetraOwnedMap );
Tpetra::Vector<double,int> tpetraOverlapVector (tpetraOverlapMap);
// The test is as follows: initialize the owned and overlap vectors
// with global IDs in the owned regions. Initialize the overlap
// vectors to equal -1 in the overlap regions. Then perform a
// communication from the owned vectors to the overlap vectors. The
// resulting overlap vectors should have global IDs everywhere and
// all of the -1 values should be overwritten.
// Initialize. We cannot assign directly to the Tpetra Vectors;
// instead, we extract nonconst views and assign to those. The
// results aren't guaranteed to be committed to the vector unless
// the views are released (by assigning Teuchos::null to them).
epetraOverlapVector.PutScalar(-1);
tpetraOverlapVector.putScalar(-1);
ArrayRCP<double> tpetraOwnedArray = tpetraOwnedVector.getDataNonConst(0);
ArrayRCP<double> tpetraOverlapArray = tpetraOverlapVector.getDataNonConst(0);
for (int owned_lid = 0;
owned_lid < tpetraOwnedMap->getNodeElementList().size();
++owned_lid) {
int gid = tpetraOwnedMap->getGlobalElement(owned_lid);
int overlap_lid = tpetraOverlapMap->getLocalElement(gid);
epetraOwnedVector[owned_lid] = gid;
epetraOverlapVector[overlap_lid] = gid;
tpetraOwnedArray[owned_lid] = gid;
tpetraOverlapArray[overlap_lid] = gid;
}
// Make sure that the changes to the Tpetra Vector were committed,
// by releasing the nonconst views.
tpetraOwnedArray = Teuchos::null;
tpetraOverlapArray = Teuchos::null;
// Test the Epetra and Tpetra Import.
if (verbose) {
out << "Testing Import from owned Map to overlap Map:" << endl << endl;
}
epetraOverlapVector.Import( epetraOwnedVector, epetraImporter, Insert);
tpetraOverlapVector.doImport(tpetraOwnedVector, tpetraImporter,
Tpetra::INSERT);
// Check the Import results.
success = countFailures (teuchosComm, epetraOwnedMap, epetraOwnedVector,
epetraOverlapMap, epetraOverlapVector,
tpetraOwnedMap, tpetraOwnedVector,
tpetraOverlapMap, tpetraOverlapVector, verbose);
const bool testOtherDirections = false;
if (testOtherDirections) {
//
// Reinitialize the Tpetra vectors and test whether Export works.
//
tpetraOverlapVector.putScalar(-1);
tpetraOwnedArray = tpetraOwnedVector.getDataNonConst(0);
tpetraOverlapArray = tpetraOverlapVector.getDataNonConst(0);
for (int owned_lid = 0;
owned_lid < tpetraOwnedMap->getNodeElementList().size();
++owned_lid)
{
int gid = tpetraOwnedMap->getGlobalElement(owned_lid);
int overlap_lid = tpetraOverlapMap->getLocalElement(gid);
tpetraOwnedArray[owned_lid] = gid;
tpetraOverlapArray[overlap_lid] = gid;
}
// Make sure that the changes to the Tpetra Vector were committed,
// by releasing the nonconst views.
tpetraOwnedArray = Teuchos::null;
tpetraOverlapArray = Teuchos::null;
// Make a Tpetra Export object, and test the export.
Tpetra::Export<int> tpetraExporter1 (tpetraOwnedMap, tpetraOverlapMap);
if (verbose) {
out << "Testing Export from owned Map to overlap Map:" << endl << endl;
}
tpetraOverlapVector.doExport (tpetraOwnedVector, tpetraExporter1,
Tpetra::INSERT);
// Check the Export results.
success = countFailures (teuchosComm, epetraOwnedMap, epetraOwnedVector,
epetraOverlapMap, epetraOverlapVector,
tpetraOwnedMap, tpetraOwnedVector,
tpetraOverlapMap, tpetraOverlapVector, verbose);
//
// Reinitialize the Tpetra vectors and see what Import in the
// other direction does.
//
tpetraOverlapVector.putScalar(-1);
tpetraOwnedArray = tpetraOwnedVector.getDataNonConst(0);
tpetraOverlapArray = tpetraOverlapVector.getDataNonConst(0);
for (int owned_lid = 0;
owned_lid < tpetraOwnedMap->getNodeElementList().size();
++owned_lid)
{
int gid = tpetraOwnedMap->getGlobalElement(owned_lid);
int overlap_lid = tpetraOverlapMap->getLocalElement(gid);
tpetraOwnedArray[owned_lid] = gid;
tpetraOverlapArray[overlap_lid] = gid;
}
// Make sure that the changes to the Tpetra Vector were committed,
// by releasing the nonconst views.
tpetraOwnedArray = Teuchos::null;
tpetraOverlapArray = Teuchos::null;
if (verbose) {
out << "Testing Import from overlap Map to owned Map:" << endl << endl;
}
Tpetra::Import<int> tpetraImporter2 (tpetraOverlapMap, tpetraOwnedMap);
tpetraOwnedVector.doImport (tpetraOverlapVector, tpetraImporter2,
Tpetra::INSERT);
// Check the Import results.
success = countFailures (teuchosComm, epetraOwnedMap, epetraOwnedVector,
epetraOverlapMap, epetraOverlapVector,
tpetraOwnedMap, tpetraOwnedVector,
tpetraOverlapMap, tpetraOverlapVector, verbose);
} // if testOtherDirections
out << "End Result: TEST " << (success ? "PASSED" : "FAILED") << endl;
MPI_Finalize ();
return success ? EXIT_SUCCESS : EXIT_FAILURE;
}
//
// TestMrhsSolver.cpp reads in a matrix in Harwell-Boeing format,
// calls one of the sparse direct solvers, using multiple right hand sides
// (one per solve) and computes the error and residual.
//
// TestSolver ignores the Harwell-Boeing right hand sides, creating
// random right hand sides instead.
//
// TestMrhsSolver can test either A x = b or A^T x = b.
// This can be a bit confusing because sparse direct solvers
// use compressed column storage - the transpose of Trilinos'
// sparse row storage.
//
// Matrices:
// readA - Serial. As read from the file.
// transposeA - Serial. The transpose of readA.
// serialA - if (transpose) then transposeA else readA
// distributedA - readA distributed to all processes
// passA - if ( distributed ) then distributedA else serialA
//
//
int Amesos_TestMrhsSolver( Epetra_Comm &Comm, char *matrix_file, int numsolves,
SparseSolverType SparseSolver, bool transpose,
int special, AMESOS_MatrixType matrix_type ) {
Comm.Barrier();
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
std::string FileName = matrix_file ;
int FN_Size = FileName.size() ;
std::string LastFiveBytes = FileName.substr( EPETRA_MAX(0,FN_Size-5), FN_Size );
std::string LastFourBytes = FileName.substr( EPETRA_MAX(0,FN_Size-4), FN_Size );
bool NonContiguousMap = false;
if ( LastFiveBytes == ".triU" ) {
// Call routine to read in unsymmetric Triplet matrix
NonContiguousMap = true;
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, false, Comm, readMap, readA, readx,
readb, readxexact, NonContiguousMap ) );
} else {
if ( LastFiveBytes == ".triS" ) {
NonContiguousMap = true;
// Call routine to read in symmetric Triplet matrix
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, true, Comm, readMap, readA, readx,
readb, readxexact) );
} else {
if ( LastFourBytes == ".mtx" ) {
EPETRA_CHK_ERR( Trilinos_Util_ReadMatrixMarket2Epetra( matrix_file, Comm, readMap,
readA, readx, readb, readxexact) );
} else {
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra( matrix_file, Comm, readMap, readA, readx,
readb, readxexact) ;
}
}
}
Epetra_CrsMatrix transposeA(Copy, *readMap, 0);
Epetra_CrsMatrix *serialA ;
if ( transpose ) {
assert( CrsMatrixTranspose( readA, &transposeA ) == 0 );
serialA = &transposeA ;
} else {
serialA = readA ;
}
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
Epetra_Map* map_;
if( NonContiguousMap ) {
//
// map gives us NumMyElements and MyFirstElement;
//
int NumGlobalElements = readMap->NumGlobalElements();
int NumMyElements = map.NumMyElements();
int MyFirstElement = map.MinMyGID();
std::vector<int> MapMap_( NumGlobalElements );
readMap->MyGlobalElements( &MapMap_[0] ) ;
Comm.Broadcast( &MapMap_[0], NumGlobalElements, 0 ) ;
map_ = new Epetra_Map( NumGlobalElements, NumMyElements, &MapMap_[MyFirstElement], 0, Comm);
} else {
map_ = new Epetra_Map( map ) ;
}
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, *map_);
Epetra_CrsMatrix A(Copy, *map_, 0);
Epetra_RowMatrix * passA = 0;
Epetra_MultiVector * passx = 0;
Epetra_MultiVector * passb = 0;
Epetra_MultiVector * passxexact = 0;
Epetra_MultiVector * passresid = 0;
Epetra_MultiVector * passtmp = 0;
Epetra_MultiVector x(*map_,numsolves);
Epetra_MultiVector b(*map_,numsolves);
Epetra_MultiVector xexact(*map_,numsolves);
Epetra_MultiVector resid(*map_,numsolves);
Epetra_MultiVector tmp(*map_,numsolves);
Epetra_MultiVector serialx(*readMap,numsolves);
Epetra_MultiVector serialb(*readMap,numsolves);
Epetra_MultiVector serialxexact(*readMap,numsolves);
Epetra_MultiVector serialresid(*readMap,numsolves);
Epetra_MultiVector serialtmp(*readMap,numsolves);
bool distribute_matrix = ( matrix_type == AMESOS_Distributed ) ;
if ( distribute_matrix ) {
//
// Initialize x, b and xexact to the values read in from the file
//
A.Export(*serialA, exporter, Add);
Comm.Barrier();
assert(A.FillComplete()==0);
Comm.Barrier();
passA = &A;
passx = &x;
passb = &b;
passxexact = &xexact;
passresid = &resid;
passtmp = &tmp;
} else {
passA = serialA;
passx = &serialx;
passb = &serialb;
passxexact = &serialxexact;
passresid = &serialresid;
passtmp = &serialtmp;
}
passxexact->SetSeed(131) ;
passxexact->Random();
passx->SetSeed(11231) ;
passx->Random();
passb->PutScalar( 0.0 );
passA->Multiply( transpose, *passxexact, *passb ) ;
Epetra_MultiVector CopyB( *passb ) ;
double Anorm = passA->NormInf() ;
SparseDirectTimingVars::SS_Result.Set_Anorm(Anorm) ;
Epetra_LinearProblem Problem( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb );
double max_resid = 0.0;
for ( int j = 0 ; j < special+1 ; j++ ) {
Epetra_Time TotalTime( Comm ) ;
if ( false ) {
#ifdef TEST_UMFPACK
unused code
} else if ( SparseSolver == UMFPACK ) {
UmfpackOO umfpack( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
umfpack.SetTrans( transpose ) ;
umfpack.Solve() ;
#endif
#ifdef TEST_SUPERLU
} else if ( SparseSolver == SuperLU ) {
SuperluserialOO superluserial ;
superluserial.SetUserMatrix( (Epetra_RowMatrix *) passA) ;
superluserial.SetPermc( SuperLU_permc ) ;
superluserial.SetTrans( transpose ) ;
superluserial.SetUseDGSSV( special == 0 ) ;
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
superluserial.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
superluserial.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
// superluserial.SetRHS( (Epetra_MultiVector *) passb_i ;
superluserial.Solve() ;
if ( i == 0 ) {
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
} else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_SLUD
} else if ( SparseSolver == SuperLUdist ) {
SuperludistOO superludist( Problem ) ;
superludist.SetTrans( transpose ) ;
bool factor = true;
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( superludist.Solve( factor ) );
factor = false;
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_SLUD2
} else if ( SparseSolver == SuperLUdist2 ) {
Superludist2_OO superludist2( Problem ) ;
superludist2.SetTrans( transpose ) ;
bool factor = true;
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( superludist2.Solve( factor ) );
factor = false;
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_DSCPACK
} else if ( SparseSolver == DSCPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Dscpack dscpack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( dscpack.SetParameters( ParamList ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( dscpack.Solve( ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_UMFPACK
} else if ( SparseSolver == UMFPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Umfpack umfpack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( umfpack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( umfpack.SetUseTranspose( transpose ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( umfpack.Solve( ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_SUPERLU
} else if ( SparseSolver == SUPERLU ) {
Teuchos::ParameterList ParamList ;
Amesos_Superlu superlu( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( superlu.SetParameters( ParamList ) );
EPETRA_CHK_ERR( superlu.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( superlu.SymbolicFactorization( ) );
EPETRA_CHK_ERR( superlu.NumericFactorization( ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( superlu.Solve( ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_SLUS
} else if ( SparseSolver == SuperLU ) {
Epetra_SLU superluserial( &Problem ) ;
bool factor = true;
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( superluserial.Solve( true, false, factor, 2, -1, true, transpose ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_KLU
} else if ( SparseSolver == KLU ) {
Teuchos::ParameterList ParamList ;
// ParamList.set("OutputLevel",2);
Amesos_Klu klu( Problem ) ;
// ParamList.set ("ScaleMethod", 0) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( klu.SetParameters( ParamList ) );
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( klu.SetParameters( ParamList ) );
EPETRA_CHK_ERR( klu.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( klu.SymbolicFactorization( ) );
for ( int trials = 0 ; trials <= 1 ; trials++) {
EPETRA_CHK_ERR( klu.NumericFactorization( ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( klu.Solve( ) );
if ( i == 0 ) {
SparseDirectTimingVars::SS_Result.Set_First_Time(
TotalTime.ElapsedTime() );
} else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time(
TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time(
TotalTime.ElapsedTime() );
}
}
}
#endif
#ifdef HAVE_AMESOS_LAPACK
} else if ( SparseSolver == LAPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Lapack lapack( Problem ) ;
EPETRA_CHK_ERR( lapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( lapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( lapack.NumericFactorization( ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( lapack.Solve( ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_TAUCS
} else if ( SparseSolver == TAUCS ) {
Teuchos::ParameterList ParamList ;
Amesos_Taucs taucs( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( taucs.SetParameters( ParamList ) );
EPETRA_CHK_ERR( taucs.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( taucs.SymbolicFactorization( ) );
EPETRA_CHK_ERR( taucs.NumericFactorization( ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( taucs.Solve( ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_PARDISO
} else if ( SparseSolver == PARDISO ) {
Teuchos::ParameterList ParamList ;
Amesos_Pardiso pardiso( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( pardiso.SetParameters( ParamList ) );
EPETRA_CHK_ERR( pardiso.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( pardiso.SymbolicFactorization( ) );
EPETRA_CHK_ERR( pardiso.NumericFactorization( ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( pardiso.Solve( ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_PARAKLETE
} else if ( SparseSolver == PARAKLETE ) {
Teuchos::ParameterList ParamList ;
Amesos_Paraklete paraklete( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( paraklete.SetParameters( ParamList ) );
EPETRA_CHK_ERR( paraklete.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( paraklete.SymbolicFactorization( ) );
EPETRA_CHK_ERR( paraklete.NumericFactorization( ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( paraklete.Solve( ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#if defined(HAVE_AMESOS_MUMPS) && defined(HAVE_MPI)
} else if ( SparseSolver == MUMPS ) {
Teuchos::ParameterList ParamList ;
Amesos_Mumps mumps( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( mumps.SetParameters( ParamList ) );
EPETRA_CHK_ERR( mumps.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( mumps.SymbolicFactorization( ) );
EPETRA_CHK_ERR( mumps.NumericFactorization( ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( mumps.Solve( ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_SCALAPACK
} else if ( SparseSolver == SCALAPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Scalapack scalapack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( scalapack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( scalapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( scalapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( scalapack.NumericFactorization( ) );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( scalapack.Solve( ) );
if ( i == 0 )
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
else {
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
}
#endif
#ifdef HAVE_AMESOS_SUPERLUDIST
} else if ( SparseSolver == SUPERLUDIST ) {
Teuchos::ParameterList ParamList ;
ParamList.set( "MaxProcs", -3 );
Amesos_Superludist superludist( Problem ) ;
EPETRA_CHK_ERR( superludist.SetParameters( ParamList ) );
EPETRA_CHK_ERR( superludist.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( superludist.SymbolicFactorization( ) );
EPETRA_CHK_ERR( superludist.NumericFactorization( ) );
SparseDirectTimingVars::SS_Result.Set_First_Time( TotalTime.ElapsedTime() );
for ( int i= 0 ; i < numsolves ; i++ ) {
// set up to sovle A X[:,i] = B[:,i]
Epetra_Vector *passb_i = (*passb)(i) ;
Epetra_Vector *passx_i = (*passx)(i) ;
Problem.SetLHS( dynamic_cast<Epetra_MultiVector *>(passx_i) ) ;
Problem.SetRHS( dynamic_cast<Epetra_MultiVector *>(passb_i) );
EPETRA_CHK_ERR( superludist.Solve( ) );
if ( i < numsolves-1 )
SparseDirectTimingVars::SS_Result.Set_Middle_Time( TotalTime.ElapsedTime() );
else
SparseDirectTimingVars::SS_Result.Set_Last_Time( TotalTime.ElapsedTime() );
}
#endif
#ifdef TEST_SPOOLES
} else if ( SparseSolver == SPOOLES ) {
SpoolesOO spooles( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
spooles.SetTrans( transpose ) ;
spooles.Solve() ;
#endif
#ifdef TEST_SPOOLESSERIAL
} else if ( SparseSolver == SPOOLESSERIAL ) {
SpoolesserialOO spoolesserial( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb ) ;
spoolesserial.Solve() ;
#endif
} else {
SparseDirectTimingVars::log_file << "Solver not implemented yet" << std::endl ;
std::cerr << "\n\n#################### Requested solver not available (Or not tested with multiple RHS) on this platform #####################\n" << std::endl ;
}
SparseDirectTimingVars::SS_Result.Set_Total_Time( TotalTime.ElapsedTime() );
//
// Compute the error = norm(xcomp - xexact )
//
std::vector <double> error(numsolves) ;
double max_error = 0.0;
passresid->Update(1.0, *passx, -1.0, *passxexact, 0.0);
passresid->Norm2(&error[0]);
for ( int i = 0 ; i< numsolves; i++ )
if ( error[i] > max_error ) max_error = error[i] ;
SparseDirectTimingVars::SS_Result.Set_Error(max_error) ;
// passxexact->Norm2(&error[0] ) ;
// passx->Norm2(&error ) ;
//
// Compute the residual = norm(Ax - b)
//
std::vector <double> residual(numsolves) ;
passtmp->PutScalar(0.0);
passA->Multiply( transpose, *passx, *passtmp);
passresid->Update(1.0, *passtmp, -1.0, *passb, 0.0);
// passresid->Update(1.0, *passtmp, -1.0, CopyB, 0.0);
passresid->Norm2(&residual[0]);
for ( int i = 0 ; i< numsolves; i++ )
if ( residual[i] > max_resid ) max_resid = residual[i] ;
SparseDirectTimingVars::SS_Result.Set_Residual(max_resid) ;
std::vector <double> bnorm(numsolves);
passb->Norm2( &bnorm[0] ) ;
SparseDirectTimingVars::SS_Result.Set_Bnorm(bnorm[0]) ;
std::vector <double> xnorm(numsolves);
passx->Norm2( &xnorm[0] ) ;
SparseDirectTimingVars::SS_Result.Set_Xnorm(xnorm[0]) ;
}
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
delete map_;
Comm.Barrier();
return 0;
}
int main(int argc, char *argv[]) {
#ifdef HAVE_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
int MyPID = Comm.MyPID();
// matrix downloaded from MatrixMarket
char FileName[] = "../HBMatrices/fidap005.rua";
Epetra_Map * readMap; // Pointers because of Trilinos_Util_ReadHb2Epetra
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra(FileName, Comm, readMap, readA, readx,
readb, readxexact);
int NumGlobalElements = readMap->NumGlobalElements();
// Create uniform distributed map
Epetra_Map map(NumGlobalElements, 0, Comm);
// Create Exporter to distribute read-in matrix and vectors
Epetra_Export exporter(*readMap, map);
Epetra_CrsMatrix A(Copy, map, 0);
Epetra_Vector x(map);
Epetra_Vector b(map);
Epetra_Vector xexact(map);
Epetra_Time FillTimer(Comm);
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
double vectorRedistributeTime = FillTimer.ElapsedTime();
A.Export(*readA, exporter, Add);
Comm.Barrier();
double matrixRedistributeTime = FillTimer.ElapsedTime() - vectorRedistributeTime;
A.FillComplete();
Comm.Barrier();
double fillCompleteTime = FillTimer.ElapsedTime() - matrixRedistributeTime;
if( MyPID==0 ) {
cout << "Vector redistribute time (sec) = "
<< vectorRedistributeTime<< endl;
cout << "Matrix redistribute time (sec) = "
<< matrixRedistributeTime << endl;
cout << "Transform to Local time (sec) = "
<< fillCompleteTime << endl<< endl;
}
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
return(EXIT_SUCCESS);
}
//
// Diagonal: 0=no change, 1=eliminate entry
// from the map for the largest row element in process 0
// 2=add diagonal entries to the matrix, with a zero value
// (assume row map contains all diagonal entries).
//
// ReindexRowMap:
// 0=no change, 1= add 2 (still contiguous), 2=non-contiguous
//
// ReindexColMap
// 0=same as RowMap, 1=add 4 - Different From RowMap, but contiguous)
//
// RangeMap:
// 0=no change, 1=serial map, 2=bizarre distribution, 3=replicated map
//
// DomainMap:
// 0=no change, 1=serial map, 2=bizarre distribution, 3=replicated map
//
RCP<Epetra_CrsMatrix> NewMatNewMap(Epetra_CrsMatrix& In,
int Diagonal,
int ReindexRowMap,
int ReindexColMap,
int RangeMapType,
int DomainMapType
)
{
//
// If we are making no change, return the original matrix (which has a linear map)
//
#if 0
std::cout << __FILE__ << "::" << __LINE__ << " "
<< Diagonal << " "
<< ReindexRowMap << " "
<< ReindexColMap << " "
<< RangeMapType << " "
<< DomainMapType << " " << std::endl ;
#endif
if ( Diagonal + ReindexRowMap + ReindexColMap + RangeMapType + DomainMapType == 0 ) {
RCP<Epetra_CrsMatrix> ReturnOrig = rcp( &In, false );
return ReturnOrig ;
}
//
// Diagonal==2 is used for a different purpose -
// Making sure that the diagonal of the matrix is non-empty.
// Note: The diagonal must exist in In.RowMap().
//
if ( Diagonal == 2 ) {
assert( ReindexRowMap==0 && ReindexColMap == 0 ) ;
}
int (*RowPermute)(int in) = 0;
int (*ColPermute)(int in) = 0;
assert( Diagonal >= 0 && Diagonal <= 2 );
assert( ReindexRowMap>=0 && ReindexRowMap<=2 );
assert( ReindexColMap>=0 && ReindexColMap<=1 );
assert( RangeMapType>=0 && RangeMapType<=3 );
assert( DomainMapType>=0 && DomainMapType<=3 );
Epetra_Map DomainMap = In.DomainMap();
Epetra_Map RangeMap = In.RangeMap();
Epetra_Map ColMap = In.ColMap();
Epetra_Map RowMap = In.RowMap();
int NumMyRowElements = RowMap.NumMyElements();
int NumMyColElements = ColMap.NumMyElements();
int NumMyRangeElements = RangeMap.NumMyElements();
int NumMyDomainElements = DomainMap.NumMyElements();
int NumGlobalRowElements = RowMap.NumGlobalElements();
int NumGlobalColElements = ColMap.NumGlobalElements();
int NumGlobalRangeElements = RangeMap.NumGlobalElements();
int NumGlobalDomainElements = DomainMap.NumGlobalElements();
assert( NumGlobalRangeElements == NumGlobalDomainElements ) ;
std::vector<int> MyGlobalRowElements( NumMyRowElements ) ;
std::vector<int> NumEntriesPerRow( NumMyRowElements ) ;
std::vector<int> MyPermutedGlobalRowElements( NumMyRowElements ) ;
std::vector<int> MyGlobalColElements( NumMyColElements ) ;
std::vector<int> MyPermutedGlobalColElements( NumMyColElements ) ; // Used to create the column map
std::vector<int> MyPermutedGlobalColElementTable( NumMyColElements ) ; // To convert local indices to global
std::vector<int> MyGlobalRangeElements( NumMyRangeElements ) ;
std::vector<int> MyPermutedGlobalRangeElements( NumMyRangeElements ) ;
std::vector<int> MyGlobalDomainElements( NumMyDomainElements ) ;
std::vector<int> MyPermutedGlobalDomainElements( NumMyDomainElements ) ;
RowMap.MyGlobalElements(&MyGlobalRowElements[0]);
ColMap.MyGlobalElements(&MyGlobalColElements[0]);
RangeMap.MyGlobalElements(&MyGlobalRangeElements[0]);
DomainMap.MyGlobalElements(&MyGlobalDomainElements[0]);
switch( ReindexRowMap ) {
case 0:
RowPermute = &NoPermute ;
break;
case 1:
RowPermute = &SmallRowPermute ;
break;
case 2:
RowPermute = BigRowPermute ;
break;
}
switch( ReindexColMap ) {
case 0:
ColPermute = RowPermute ;
break;
case 1:
ColPermute = &SmallColPermute ;
break;
}
//
// Create Serial Range and Domain Maps based on the permuted indexing
//
int nlocal = 0;
if (In.Comm().MyPID()==0) nlocal = NumGlobalRangeElements;
std::vector<int> AllIDs( NumGlobalRangeElements ) ;
for ( int i = 0; i < NumGlobalRangeElements ; i++ ) AllIDs[i] = (*RowPermute)( i ) ;
Epetra_Map SerialRangeMap( -1, nlocal, &AllIDs[0], 0, In.Comm());
std::vector<int> AllIDBs( NumGlobalRangeElements ) ;
for ( int i = 0; i < NumGlobalRangeElements ; i++ ) AllIDBs[i] = (*ColPermute)( i ) ;
Epetra_Map SerialDomainMap( -1, nlocal, &AllIDBs[0], 0, In.Comm());
//
// Create Bizarre Range and Domain Maps based on the permuted indexing
// These are nearly serial, having all but one element on process 0
// The goal here is to make sure that we can use Domain and Range maps
// that are neither serial, nor distributed in the normal manner.
//
std::vector<int> AllIDCs( NumGlobalRangeElements ) ;
for ( int i = 0; i < NumGlobalRangeElements ; i++ ) AllIDCs[i] = (*ColPermute)( i ) ;
if ( In.Comm().NumProc() > 1 ) {
if (In.Comm().MyPID()==0) nlocal = NumGlobalRangeElements-1;
if (In.Comm().MyPID()==1) {
nlocal = 1;
AllIDCs[0] = (*ColPermute)( NumGlobalRangeElements - 1 );
}
}
int iam = In.Comm().MyPID();
Epetra_Map BizarreDomainMap( -1, nlocal, &AllIDCs[0], 0, In.Comm());
std::vector<int> AllIDDs( NumGlobalRangeElements ) ;
for ( int i = 0; i < NumGlobalRangeElements ; i++ ) AllIDDs[i] = (*RowPermute)( i ) ;
if ( In.Comm().NumProc() > 1 ) {
if (In.Comm().MyPID()==0) nlocal = NumGlobalRangeElements-1;
if (In.Comm().MyPID()==1) {
nlocal = 1;
AllIDDs[0] = (*RowPermute)( NumGlobalRangeElements -1 ) ;
}
}
Epetra_Map BizarreRangeMap( -1, nlocal, &AllIDDs[0], 0, In.Comm());
//
// Compute the column map
//
// If Diagonal==1, remove the column corresponding to the last row owned
// by process 0. Removing this column from a tridiagonal matrix, leaves
// a disconnected, but non-singular matrix.
//
int NumMyColElementsOut = 0 ;
int NumGlobalColElementsOut ;
if ( Diagonal == 1 )
NumGlobalColElementsOut = NumGlobalColElements-1;
else
NumGlobalColElementsOut = NumGlobalColElements;
if ( Diagonal == 1 && iam==0 ) {
for ( int i=0; i < NumMyColElements ; i++ ) {
if ( MyGlobalColElements[i] != MyGlobalRowElements[NumMyRowElements-1] ) {
MyPermutedGlobalColElements[NumMyColElementsOut++] =
(*ColPermute)( MyGlobalColElements[i] ) ;
}
}
assert( NumMyColElementsOut == NumMyColElements-1 );
} else {
for ( int i=0; i < NumMyColElements ; i++ )
MyPermutedGlobalColElements[i] =
(*ColPermute)( MyGlobalColElements[i] ) ;
NumMyColElementsOut = NumMyColElements ;
if ( Diagonal == 2 ) {
// For each row, make sure that the column map has this row in it,
// if it doesn't, add it to the column map.
// Note: MyPermutedGlobalColElements == MyGlobalColElements when
// Diagonal==2 because ( Diagonal == 2 ) implies:
// ReindexRowMap==0 && ReindexColMap == 0 - see assert above
for ( int i=0; i < NumMyRowElements ; i++ ) {
bool MissingDiagonal = true;
for ( int j=0; j < NumMyColElements; j++ ) {
if ( MyGlobalRowElements[i] == MyGlobalColElements[j] ) {
MissingDiagonal = false;
}
}
if ( MissingDiagonal ) {
MyPermutedGlobalColElements.resize(NumMyColElements+1);
MyPermutedGlobalColElements[NumMyColElementsOut] = MyGlobalRowElements[i];
NumMyColElementsOut++;
}
}
In.Comm().SumAll(&NumMyColElementsOut,&NumGlobalColElementsOut,1);
}
}
//
// These tables are used both as the permutation tables and to create the maps.
//
for ( int i=0; i < NumMyColElements ; i++ )
MyPermutedGlobalColElementTable[i] =
(*ColPermute)( MyGlobalColElements[i] ) ;
for ( int i=0; i < NumMyRowElements ; i++ )
MyPermutedGlobalRowElements[i] =
(*RowPermute)( MyGlobalRowElements[i] ) ;
for ( int i=0; i < NumMyRangeElements ; i++ )
MyPermutedGlobalRangeElements[i] =
(*RowPermute)( MyGlobalRangeElements[i] ) ;
for ( int i=0; i < NumMyDomainElements ; i++ )
MyPermutedGlobalDomainElements[i] =
(*ColPermute)( MyGlobalDomainElements[i] ) ;
RCP<Epetra_Map> PermutedRowMap =
rcp( new Epetra_Map( NumGlobalRowElements, NumMyRowElements,
&MyPermutedGlobalRowElements[0], 0, In.Comm() ) );
RCP<Epetra_Map> PermutedColMap =
rcp( new Epetra_Map( NumGlobalColElementsOut, NumMyColElementsOut,
&MyPermutedGlobalColElements[0], 0, In.Comm() ) );
RCP<Epetra_Map> PermutedRangeMap =
rcp( new Epetra_Map( NumGlobalRangeElements, NumMyRangeElements,
&MyPermutedGlobalRangeElements[0], 0, In.Comm() ) );
RCP<Epetra_Map> PermutedDomainMap =
rcp( new Epetra_Map( NumGlobalDomainElements, NumMyDomainElements,
&MyPermutedGlobalDomainElements[0], 0, In.Comm() ) );
//
// These vectors are filled and then passed to InsertGlobalValues
//
std::vector<int> ThisRowIndices( In.MaxNumEntries() );
std::vector<double> ThisRowValues( In.MaxNumEntries() );
std::vector<int> PermutedGlobalColIndices( In.MaxNumEntries() );
//std::cout << __FILE__ << "::" <<__LINE__ << std::endl ;
RCP<Epetra_CrsMatrix> Out =
rcp( new Epetra_CrsMatrix( Copy, *PermutedRowMap, *PermutedColMap, 0 ) );
for (int i=0; i<NumMyRowElements; i++)
{
int NumIndicesThisRow = 0;
assert( In.ExtractMyRowCopy( i,
In.MaxNumEntries(),
NumIndicesThisRow,
&ThisRowValues[0],
&ThisRowIndices[0] ) == 0 ) ;
for (int j = 0 ; j < NumIndicesThisRow ; j++ )
{
PermutedGlobalColIndices[j] = MyPermutedGlobalColElementTable[ ThisRowIndices[j] ] ;
}
bool MissingDiagonal = false;
if ( Diagonal==2 ) {
//
assert( MyGlobalRowElements[i] == MyPermutedGlobalRowElements[i] );
MissingDiagonal = true;
for( int j =0 ; j < NumIndicesThisRow ; j++ ) {
if ( PermutedGlobalColIndices[j] == MyPermutedGlobalRowElements[i] ) {
MissingDiagonal = false ;
}
}
#if 0
std::cout << __FILE__ << "::" << __LINE__
<< " i = " << i
<< " MyPermutedGlobalRowElements[i] = " << MyPermutedGlobalRowElements[i]
<< " MissingDiagonal = " << MissingDiagonal << std::endl ;
#endif
}
if ( MissingDiagonal ) {
ThisRowValues.resize(NumIndicesThisRow+1) ;
ThisRowValues[NumIndicesThisRow] = 0.0;
PermutedGlobalColIndices.resize(NumIndicesThisRow+1);
PermutedGlobalColIndices[NumIndicesThisRow] = MyPermutedGlobalRowElements[i] ;
#if 0
std::cout << __FILE__ << "::" << __LINE__
<< " i = " << i
<< "NumIndicesThisRow = " << NumIndicesThisRow
<< "ThisRowValues[NumIndicesThisRow = " << ThisRowValues[NumIndicesThisRow]
<< " PermutedGlobalColIndices[NumIndcesThisRow] = " << PermutedGlobalColIndices[NumIndicesThisRow]
<< std::endl ;
#endif
NumIndicesThisRow++ ;
}
assert( Out->InsertGlobalValues( MyPermutedGlobalRowElements[i],
NumIndicesThisRow,
&ThisRowValues[0],
&PermutedGlobalColIndices[0] ) >= 0 );
}
//
Epetra_LocalMap ReplicatedMap( NumGlobalRangeElements, 0, In.Comm() );
RCP<Epetra_Map> OutRangeMap ;
RCP<Epetra_Map> OutDomainMap ;
switch( RangeMapType ) {
case 0:
OutRangeMap = PermutedRangeMap ;
break;
case 1:
OutRangeMap = rcp(&SerialRangeMap, false);
break;
case 2:
OutRangeMap = rcp(&BizarreRangeMap, false);
break;
case 3:
OutRangeMap = rcp(&ReplicatedMap, false);
break;
}
// switch( DomainMapType ) {
switch( DomainMapType ) {
case 0:
OutDomainMap = PermutedDomainMap ;
break;
case 1:
OutDomainMap = rcp(&SerialDomainMap, false);
break;
case 2:
OutDomainMap = rcp(&BizarreDomainMap, false);
break;
case 3:
OutDomainMap = rcp(&ReplicatedMap, false);
break;
}
#if 0
assert(Out->FillComplete( *PermutedDomainMap, *PermutedRangeMap )==0);
#else
assert(Out->FillComplete( *OutDomainMap, *OutRangeMap )==0);
#endif
#if 0
std::cout << __FILE__ << "::" << __LINE__ << std::endl ;
Out->Print( std::cout ) ;
#endif
return Out;
}
int
main (int argc, char *argv[])
{
using namespace Anasazi;
using Teuchos::RCP;
using Teuchos::rcp;
using std::endl;
#ifdef HAVE_MPI
// Initialize MPI
MPI_Init (&argc, &argv);
#endif // HAVE_MPI
// Create an Epetra communicator
#ifdef HAVE_MPI
Epetra_MpiComm Comm (MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif // HAVE_MPI
// Create an Anasazi output manager
BasicOutputManager<double> printer;
printer.stream(Errors) << Anasazi_Version() << std::endl << std::endl;
// Get the sorting std::string from the command line
std::string which ("LM");
Teuchos::CommandLineProcessor cmdp (false, true);
cmdp.setOption("sort", &which, "Targetted eigenvalues (SM or LM).");
if (cmdp.parse (argc, argv) != Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL) {
#ifdef HAVE_MPI
MPI_Finalize ();
#endif // HAVE_MPI
return -1;
}
// Dimension of the matrix
//
// Discretization points in any one direction.
const int nx = 10;
// Size of matrix nx*nx
const int NumGlobalElements = nx*nx;
// Construct a Map that puts approximately the same number of
// equations on each process.
Epetra_Map Map (NumGlobalElements, 0, Comm);
// Get update list and number of local equations from newly created Map.
int NumMyElements = Map.NumMyElements ();
std::vector<int> MyGlobalElements (NumMyElements);
Map.MyGlobalElements (&MyGlobalElements[0]);
// Create an integer vector NumNz that is used to build the Petra
// matrix. NumNz[i] is the number of OFF-DIAGONAL terms for the
// i-th global equation on this process.
std::vector<int> NumNz (NumMyElements);
/* We are building a matrix of block structure:
| T -I |
|-I T -I |
| -I T |
| ... -I|
| -I T|
where each block is dimension nx by nx and the matrix is on the order of
nx*nx. The block T is a tridiagonal matrix.
*/
for (int i=0; i<NumMyElements; ++i) {
if (MyGlobalElements[i] == 0 || MyGlobalElements[i] == NumGlobalElements-1 ||
MyGlobalElements[i] == nx-1 || MyGlobalElements[i] == nx*(nx-1) ) {
NumNz[i] = 3;
}
else if (MyGlobalElements[i] < nx || MyGlobalElements[i] > nx*(nx-1) ||
MyGlobalElements[i]%nx == 0 || (MyGlobalElements[i]+1)%nx == 0) {
NumNz[i] = 4;
}
else {
NumNz[i] = 5;
}
}
// Create an Epetra_Matrix
RCP<Epetra_CrsMatrix> A = rcp (new Epetra_CrsMatrix (Copy, Map, &NumNz[0]));
// Compute coefficients for discrete convection-diffution operator
const double one = 1.0;
std::vector<double> Values(4);
std::vector<int> Indices(4);
double rho = 0.0;
double h = one /(nx+1);
double h2 = h*h;
double c = 5.0e-01*rho/ h;
Values[0] = -one/h2 - c; Values[1] = -one/h2 + c; Values[2] = -one/h2; Values[3]= -one/h2;
double diag = 4.0 / h2;
int NumEntries;
for (int i=0; i<NumMyElements; ++i) {
if (MyGlobalElements[i]==0) {
Indices[0] = 1;
Indices[1] = nx;
NumEntries = 2;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] == nx*(nx-1)) {
Indices[0] = nx*(nx-1)+1;
Indices[1] = nx*(nx-2);
NumEntries = 2;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] == nx-1) {
Indices[0] = nx-2;
NumEntries = 1;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
Indices[0] = 2*nx-1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] == NumGlobalElements-1) {
Indices[0] = NumGlobalElements-2;
NumEntries = 1;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
Indices[0] = nx*(nx-1)-1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] < nx) {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
Indices[2] = MyGlobalElements[i]+nx;
NumEntries = 3;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i] > nx*(nx-1)) {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
Indices[2] = MyGlobalElements[i]-nx;
NumEntries = 3;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
}
else if (MyGlobalElements[i]%nx == 0) {
Indices[0] = MyGlobalElements[i]+1;
Indices[1] = MyGlobalElements[i]-nx;
Indices[2] = MyGlobalElements[i]+nx;
NumEntries = 3;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[1], &Indices[0]);
assert( info==0 );
}
else if ((MyGlobalElements[i]+1)%nx == 0) {
Indices[0] = MyGlobalElements[i]-nx;
Indices[1] = MyGlobalElements[i]+nx;
NumEntries = 2;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[2], &Indices[0]);
assert( info==0 );
Indices[0] = MyGlobalElements[i]-1;
NumEntries = 1;
info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
}
else {
Indices[0] = MyGlobalElements[i]-1;
Indices[1] = MyGlobalElements[i]+1;
Indices[2] = MyGlobalElements[i]-nx;
Indices[3] = MyGlobalElements[i]+nx;
NumEntries = 4;
int info = A->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
}
// Put in the diagonal entry
int info = A->InsertGlobalValues(MyGlobalElements[i], 1, &diag, &MyGlobalElements[i]);
assert( info==0 );
}
// Finish up
int info = A->FillComplete ();
assert( info==0 );
A->SetTracebackMode (1); // Shutdown Epetra Warning tracebacks
// Create a identity matrix for the temporary mass matrix
RCP<Epetra_CrsMatrix> M = rcp (new Epetra_CrsMatrix (Copy, Map, 1));
for (int i=0; i<NumMyElements; i++) {
Values[0] = one;
Indices[0] = i;
NumEntries = 1;
info = M->InsertGlobalValues(MyGlobalElements[i], NumEntries, &Values[0], &Indices[0]);
assert( info==0 );
}
// Finish up
info = M->FillComplete ();
assert( info==0 );
M->SetTracebackMode (1); // Shutdown Epetra Warning tracebacks
//************************************
// Call the LOBPCG solver manager
//***********************************
//
// Variables used for the LOBPCG Method
const int nev = 10;
const int blockSize = 5;
const int maxIters = 500;
const double tol = 1.0e-8;
typedef Epetra_MultiVector MV;
typedef Epetra_Operator OP;
typedef MultiVecTraits<double, Epetra_MultiVector> MVT;
// Create an Epetra_MultiVector for an initial vector to start the
// solver. Note: This needs to have the same number of columns as
// the blocksize.
RCP<Epetra_MultiVector> ivec = rcp (new Epetra_MultiVector (Map, blockSize));
ivec->Random (); // fill the initial vector with random values
// Create the eigenproblem.
RCP<BasicEigenproblem<double, MV, OP> > MyProblem =
rcp (new BasicEigenproblem<double, MV, OP> (A, ivec));
// Inform the eigenproblem that the operator A is symmetric
MyProblem->setHermitian (true);
// Set the number of eigenvalues requested
MyProblem->setNEV (nev);
// Tell the eigenproblem that you are finishing passing it information.
const bool success = MyProblem->setProblem ();
if (! success) {
printer.print (Errors, "Anasazi::BasicEigenproblem::setProblem() reported an error.\n");
#ifdef HAVE_MPI
MPI_Finalize ();
#endif // HAVE_MPI
return -1;
}
// Create parameter list to pass into the solver manager
Teuchos::ParameterList MyPL;
MyPL.set ("Which", which);
MyPL.set ("Block Size", blockSize);
MyPL.set ("Maximum Iterations", maxIters);
MyPL.set ("Convergence Tolerance", tol);
MyPL.set ("Full Ortho", true);
MyPL.set ("Use Locking", true);
// Create the solver manager
LOBPCGSolMgr<double, MV, OP> MySolverMan (MyProblem, MyPL);
// Solve the problem
ReturnType returnCode = MySolverMan.solve ();
// Get the eigenvalues and eigenvectors from the eigenproblem
Eigensolution<double,MV> sol = MyProblem->getSolution ();
std::vector<Value<double> > evals = sol.Evals;
RCP<MV> evecs = sol.Evecs;
// Compute residuals.
std::vector<double> normR (sol.numVecs);
if (sol.numVecs > 0) {
Teuchos::SerialDenseMatrix<int,double> T (sol.numVecs, sol.numVecs);
Epetra_MultiVector tempAevec (Map, sol.numVecs );
T.putScalar (0.0);
for (int i = 0; i < sol.numVecs; ++i) {
T(i,i) = evals[i].realpart;
}
A->Apply (*evecs, tempAevec);
MVT::MvTimesMatAddMv (-1.0, *evecs, T, 1.0, tempAevec);
MVT::MvNorm (tempAevec, normR);
}
// Print the results
std::ostringstream os;
os.setf (std::ios_base::right, std::ios_base::adjustfield);
os << "Solver manager returned "
<< (returnCode == Converged ? "converged." : "unconverged.") << endl;
os << endl;
os << "------------------------------------------------------" << endl;
os << std::setw(16) << "Eigenvalue"
<< std::setw(18) << "Direct Residual"
<< endl;
os << "------------------------------------------------------" << endl;
for (int i = 0; i < sol.numVecs; ++i) {
os << std::setw(16) << evals[i].realpart
<< std::setw(18) << normR[i] / evals[i].realpart
<< endl;
}
os << "------------------------------------------------------" << endl;
printer.print (Errors, os.str ());
#ifdef HAVE_MPI
MPI_Finalize ();
#endif // HAVE_MPI
return 0;
}
void build_simple_matrix(
Epetra_Comm &comm, // Communicator to use
Epetra_CrsMatrix *&A, // OUTPUT: Matrix returned
itype nGlobalRows, // Number of global matrix rows and columns
bool testEpetra64, // if true, add 2*INT_MAX to each global ID
// to exercise Epetra64
bool verbose // if true, print out matrix information
)
{
Epetra_Map *rowMap = NULL; // Row map for the created matrix
Epetra_Map *colMap = NULL; // Col map for the created matrix
Epetra_Map *vectorMap = NULL; // Range/Domain map for the created matrix
long long offsetEpetra64;
build_maps(nGlobalRows, testEpetra64, comm,
&vectorMap, &rowMap, &colMap, offsetEpetra64, verbose);
// Create an integer vector nnzPerRow that is used to build the Epetra Matrix.
// nnzPerRow[i] is the number of entries for the ith global equation
int nMyRows = rowMap->NumMyElements();
std::vector<int> nnzPerRow(nMyRows+1, 0);
// Also create lists of the nonzeros to be assigned to processors.
// To save programming time and complexity, these vectors are allocated
// bigger than they may actually be needed.
std::vector<itype> iv(3*nMyRows+1);
std::vector<itype> jv(3*nMyRows+1);
std::vector<double> vv(3*nMyRows+1);
itype nMyNonzeros = 0;
for (itype i = 0, myrowcnt = 0; i < nGlobalRows; i++) {
if (rowMap->MyGID(i+offsetEpetra64)) {
// This processor owns part of this row; see whether it owns the nonzeros
if (i > 0 && (!colMap || colMap->MyGID(i-1+offsetEpetra64))) {
iv[nMyNonzeros] = i + offsetEpetra64;
jv[nMyNonzeros] = i-1 + offsetEpetra64;
vv[nMyNonzeros] = -1;
nMyNonzeros++;
nnzPerRow[myrowcnt]++;
}
if (!colMap || colMap->MyGID(i+offsetEpetra64)) {
iv[nMyNonzeros] = i + offsetEpetra64;
jv[nMyNonzeros] = i + offsetEpetra64;
vv[nMyNonzeros] = ((i == 0 || i == nGlobalRows-1) ? 1. : 2.);
nMyNonzeros++;
nnzPerRow[myrowcnt]++;
}
if (i < nGlobalRows - 1 && (!colMap || colMap->MyGID(i+1+offsetEpetra64))) {
iv[nMyNonzeros] = i + offsetEpetra64;
jv[nMyNonzeros] = i+1 + offsetEpetra64;
vv[nMyNonzeros] = -1;
nMyNonzeros++;
nnzPerRow[myrowcnt]++;
}
myrowcnt++;
}
}
// Create an Epetra_Matrix
A = new Epetra_CrsMatrix(Copy, *rowMap, &nnzPerRow[0], true);
int info;
for (int sum = 0, i=0; i < nMyRows; i++) {
if (nnzPerRow[i]) {
info = A->InsertGlobalValues(iv[sum],nnzPerRow[i],&vv[sum],&jv[sum]);
assert(info==0);
sum += nnzPerRow[i];
}
}
// Finish up
if (vectorMap)
info = A->FillComplete(*vectorMap, *vectorMap);
else
info = A->FillComplete();
assert(info==0);
}
int main(int argc, char *argv[]) {
// Initialize MPI
#ifdef HAVE_MPI
MPI_Init(&argc,&argv);
#endif
// Create a communicator for Epetra objects
#ifdef HAVE_MPI
Epetra_MpiComm Comm( MPI_COMM_WORLD );
#else
Epetra_SerialComm Comm;
#endif
bool verbose = false;
bool success = false;
try {
int globalLength = 100; // This should suffice
if (argc > 1)
if (argv[1][0]=='-' && argv[1][1]=='v')
verbose = true;
// Get the process ID and the total number of processors
int MyPID = Comm.MyPID();
// Set up the printing utilities
Teuchos::RCP<Teuchos::ParameterList> noxParamsPtr =
Teuchos::rcp(new Teuchos::ParameterList);
Teuchos::ParameterList& noxParams = *(noxParamsPtr.get());
// Only print output if the "-v" flag is set on the command line
Teuchos::ParameterList& printParams = noxParams.sublist("Printing");
printParams.set("MyPID", MyPID);
printParams.set("Output Precision", 5);
printParams.set("Output Processor", 0);
if( verbose )
printParams.set("Output Information",
NOX::Utils::OuterIteration +
NOX::Utils::OuterIterationStatusTest +
NOX::Utils::InnerIteration +
NOX::Utils::Parameters +
NOX::Utils::Details +
NOX::Utils::Warning +
NOX::Utils::TestDetails);
else
printParams.set("Output Information", NOX::Utils::Error);
NOX::Utils printing(printParams);
// Identify the test problem
if (printing.isPrintType(NOX::Utils::TestDetails))
printing.out() << "Starting epetra/NOX_Vector/NOX_Vector.exe" << std::endl;
// Create a TestCompare class
NOX::TestCompare tester( printing.out(), printing);
double tolerance = 1.e-12;
NOX::TestCompare::CompareType aComp = NOX::TestCompare::Absolute;
// Identify processor information
#ifdef HAVE_MPI
printing.out() << "Parallel Run" << std::endl;
printing.out() << "Number of processors = " << Comm.NumProc() << std::endl;
printing.out() << "Print Process = " << MyPID << std::endl;
Comm.Barrier();
if (printing.isPrintType(NOX::Utils::TestDetails))
printing.out() << "Process " << MyPID << " is alive!" << std::endl;
Comm.Barrier();
#else
printing.out() << "Serial Run" << std::endl;
#endif
// Create a map describing data distribution
Epetra_Map * standardMap = new Epetra_Map(globalLength, 0, Comm);
// Return value
int status = 0;
// *** Start Testing Here!!! ***
// First create the Epetra_Vector needed to construct our NOX vector
Epetra_Vector * epetraVec = new Epetra_Vector(*standardMap, true);
NOX::Epetra::Vector * noxVec1 = new NOX::Epetra::Vector(*epetraVec, NOX::DeepCopy);
delete epetraVec; epetraVec = 0;
NOX::Epetra::Vector * noxVec2 = new NOX::Epetra::Vector(*noxVec1);
noxVec2->init(1.0);
// Test our norms
NOX::Abstract::Vector::NormType
oneNorm = NOX::Abstract::Vector::OneNorm,
twoNorm = NOX::Abstract::Vector::TwoNorm,
infNorm = NOX::Abstract::Vector::MaxNorm;
double expectedOneNorm = (double) globalLength,
expectedTwoNorm = sqrt( (double) globalLength),
expectedInfNorm = 1.0;
status += tester.testValue( noxVec2->norm(oneNorm), expectedOneNorm,
tolerance, "One-Norm Test", aComp);
status += tester.testValue( noxVec2->norm(twoNorm), expectedTwoNorm,
tolerance, "Two-Norm Test", aComp);
status += tester.testValue( noxVec2->norm(infNorm), expectedInfNorm,
tolerance, "Max-Norm Test", aComp);
// Test random, reciprocal and dot methods
noxVec1->random();
// Threshold values since we want to do a reciprocal
int myLength = standardMap->NumMyElements();
for( int i = 0; i < myLength; ++i )
if( fabs(noxVec1->getEpetraVector()[i]) < 1.e-8 ) noxVec1->getEpetraVector()[i] = 1.e-8;
noxVec2->reciprocal(*noxVec1);
double product = noxVec1->innerProduct(*noxVec2);
status += tester.testValue( product, expectedOneNorm,
tolerance, "Random, Reciprocal and Dot Test", aComp);
// Test abs and weighted-norm methods
/* ----------------------------
NOT SUPPORTED AT THIS TIME
----------------------------
noxVec2->abs(*noxVec2);
double wNorm = noxVec1->norm(*noxVec2);
status += tester.testValue( wNorm, noxVec1->norm(oneNorm),
tolerance, "Abs and Weighted-Norm Test", aComp);
*/
// Test operator= , abs, update and scale methods
(*noxVec2) = (*noxVec1);
noxVec2->abs(*noxVec2);
double sumAll = noxVec1->norm(oneNorm);
noxVec2->update( 1.0, *noxVec1, 1.0 );
noxVec2->scale(0.5);
double sumPositive = noxVec2->norm(oneNorm);
(*noxVec2) = (*noxVec1);
noxVec2->abs(*noxVec2);
noxVec2->update( 1.0, *noxVec1, -1.0 );
noxVec2->scale(0.5);
double sumNegative = noxVec2->norm(oneNorm);
status += tester.testValue( (sumPositive + sumNegative), sumAll,
tolerance, "Abs, Operator= , Update and Scale Test", aComp);
success = status==0;
if (success)
printing.out() << "Test passed!" << std::endl;
else
printing.out() << "Test failed!" << std::endl;
delete noxVec2;
delete noxVec1;
delete standardMap;
}
TEUCHOS_STANDARD_CATCH_STATEMENTS(verbose, std::cerr, success);
#ifdef HAVE_MPI
MPI_Finalize();
#endif
return ( success ? EXIT_SUCCESS : EXIT_FAILURE );
}
int Drumm2(const Epetra_Map& map, bool verbose)
{
//Simple 2-element problem (element as in "finite-element") from
//Clif Drumm. Two triangular elements, one per processor, as shown
//here:
//
// *----*
// 3|\ 2|
// | \ |
// | 0\1|
// | \|
// *----*
// 0 1
//
//Element 0 on processor 0, element 1 on processor 1.
//Processor 0 will own nodes 0,1,3 and processor 1 will own node 2.
//Each processor will pass a 3x3 element-connectivity-matrix to
//Epetra_FECrsGraph.
//After GlobalAssemble(), the graph should be as follows:
//
// row 0: 2 1 0 1
//proc 0 row 1: 1 4 1 2
// row 2: 0 1 2 1
//----------------------------------
//proc 1 row 3: 1 2 1 4
//
int numProcs = map.Comm().NumProc();
int localProc = map.Comm().MyPID();
if (numProcs != 2) return(0);
int indexBase = 0, ierr = 0;
int numMyNodes = 3;
long long* myNodes = new long long[numMyNodes];
if (localProc == 0) {
myNodes[0] = 0;
myNodes[1] = 1;
myNodes[2] = 3;
}
else {
numMyNodes = 1;
myNodes[0] = 2;
}
Epetra_Map Map((long long) -1, numMyNodes, myNodes, indexBase, map.Comm());
int rowLengths = 3;
Epetra_FECrsGraph A(Copy, Map, rowLengths);
if (localProc != 0) {
numMyNodes = 3;
myNodes[0] = 1;
myNodes[1] = 2;
myNodes[2] = 3;
}
EPETRA_TEST_ERR( A.InsertGlobalIndices(numMyNodes, myNodes,
numMyNodes, myNodes),ierr);
EPETRA_TEST_ERR( A.GlobalAssemble(), ierr );
if (verbose) {
A.Print(std::cout);
}
delete [] myNodes;
return(0);
}
int Amesos_TestSolver( Epetra_Comm &Comm, char *matrix_file,
SparseSolverType SparseSolver,
bool transpose,
int special, AMESOS_MatrixType matrix_type ) {
Epetra_Map * readMap;
Epetra_CrsMatrix * readA;
Epetra_Vector * readx;
Epetra_Vector * readb;
Epetra_Vector * readxexact;
std::string FileName = matrix_file ;
int FN_Size = FileName.size() ;
std::string LastFiveBytes = FileName.substr( EPETRA_MAX(0,FN_Size-5), FN_Size );
std::string LastFourBytes = FileName.substr( EPETRA_MAX(0,FN_Size-4), FN_Size );
bool NonContiguousMap = false;
if ( LastFiveBytes == ".triU" ) {
// Call routine to read in unsymmetric Triplet matrix
NonContiguousMap = true;
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, false, Comm, readMap, readA, readx,
readb, readxexact, NonContiguousMap ) );
} else {
if ( LastFiveBytes == ".triS" ) {
NonContiguousMap = true;
// Call routine to read in symmetric Triplet matrix
EPETRA_CHK_ERR( Trilinos_Util_ReadTriples2Epetra( matrix_file, true, Comm, readMap, readA, readx,
readb, readxexact, NonContiguousMap ) );
} else {
if ( LastFourBytes == ".mtx" ) {
EPETRA_CHK_ERR( Trilinos_Util_ReadMatrixMarket2Epetra( matrix_file, Comm, readMap,
readA, readx, readb, readxexact) );
} else {
// Call routine to read in HB problem
Trilinos_Util_ReadHb2Epetra( matrix_file, Comm, readMap, readA, readx,
readb, readxexact) ;
}
}
}
Epetra_CrsMatrix transposeA(Copy, *readMap, 0);
Epetra_CrsMatrix *serialA ;
if ( transpose ) {
assert( CrsMatrixTranspose( readA, &transposeA ) == 0 );
serialA = &transposeA ;
} else {
serialA = readA ;
}
Epetra_RowMatrix * passA = 0;
Epetra_Vector * passx = 0;
Epetra_Vector * passb = 0;
Epetra_Vector * passxexact = 0;
Epetra_Vector * passresid = 0;
Epetra_Vector * passtmp = 0;
// Create uniform distributed map
Epetra_Map map(readMap->NumGlobalElements(), 0, Comm);
Epetra_Map* map_;
if( NonContiguousMap ) {
//
// map gives us NumMyElements and MyFirstElement;
//
int NumGlobalElements = readMap->NumGlobalElements();
int NumMyElements = map.NumMyElements();
int MyFirstElement = map.MinMyGID();
std::vector<int> MapMap_( NumGlobalElements );
readMap->MyGlobalElements( &MapMap_[0] ) ;
Comm.Broadcast( &MapMap_[0], NumGlobalElements, 0 ) ;
map_ = new Epetra_Map( NumGlobalElements, NumMyElements, &MapMap_[MyFirstElement], 0, Comm);
} else {
map_ = new Epetra_Map( map ) ;
}
Epetra_CrsMatrix A(Copy, *map_, 0);
const Epetra_Map &OriginalMap = serialA->RowMatrixRowMap() ;
assert( OriginalMap.SameAs(*readMap) );
Epetra_Export exporter(OriginalMap, *map_);
Epetra_Export exporter2(OriginalMap, *map_);
Epetra_Export MatrixExporter(OriginalMap, *map_);
Epetra_CrsMatrix AwithDiag(Copy, *map_, 0);
Epetra_Vector x(*map_);
Epetra_Vector b(*map_);
Epetra_Vector xexact(*map_);
Epetra_Vector resid(*map_);
Epetra_Vector readresid(*readMap);
Epetra_Vector tmp(*map_);
Epetra_Vector readtmp(*readMap);
// Epetra_Vector xcomp(*map_); // X as computed by the solver
bool distribute_matrix = ( matrix_type == AMESOS_Distributed ) ;
if ( distribute_matrix ) {
// Create Exporter to distribute read-in matrix and vectors
//
// Initialize x, b and xexact to the values read in from the file
//
x.Export(*readx, exporter, Add);
b.Export(*readb, exporter, Add);
xexact.Export(*readxexact, exporter, Add);
Comm.Barrier();
A.Export(*serialA, exporter, Add);
assert(A.FillComplete()==0);
Comm.Barrier();
passA = &A;
passx = &x;
passb = &b;
passxexact = &xexact;
passresid = &resid;
passtmp = &tmp;
} else {
passA = serialA;
passx = readx;
passb = readb;
passxexact = readxexact;
passresid = &readresid;
passtmp = &readtmp;
}
Epetra_MultiVector CopyB( *passb ) ;
double Anorm = passA->NormInf() ;
SparseDirectTimingVars::SS_Result.Set_Anorm(Anorm) ;
Epetra_LinearProblem Problem( (Epetra_RowMatrix *) passA,
(Epetra_MultiVector *) passx,
(Epetra_MultiVector *) passb );
for ( int i = 0; i < 1+special ; i++ ) {
Epetra_Time TotalTime( Comm ) ;
if ( false ) {
// TEST_UMFPACK is never set by configure
#ifdef HAVE_AMESOS_SUPERLUDIST
} else if ( SparseSolver == SUPERLUDIST ) {
Teuchos::ParameterList ParamList ;
ParamList.set( "MaxProcs", -3 );
Amesos_Superludist A_Superludist( Problem ) ;
//ParamList.set( "Redistribute", true );
//ParamList.set( "AddZeroToDiag", true );
Teuchos::ParameterList& SuperludistParams = ParamList.sublist("Superludist") ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_Superludist.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_Superludist.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_Superludist.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_Superludist.NumericFactorization( ) );
EPETRA_CHK_ERR( A_Superludist.Solve( ) );
#endif
#ifdef HAVE_AMESOS_DSCPACK
} else if ( SparseSolver == DSCPACK ) {
Teuchos::ParameterList ParamList ;
ParamList.set( "MaxProcs", -3 );
Amesos_Dscpack A_dscpack( Problem ) ;
EPETRA_CHK_ERR( A_dscpack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_dscpack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_dscpack.NumericFactorization( ) );
EPETRA_CHK_ERR( A_dscpack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SCALAPACK
} else if ( SparseSolver == SCALAPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Scalapack A_scalapack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_scalapack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_scalapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_scalapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_scalapack.NumericFactorization( ) );
EPETRA_CHK_ERR( A_scalapack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_TAUCS
} else if ( SparseSolver == TAUCS ) {
Teuchos::ParameterList ParamList ;
Amesos_Taucs A_taucs( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_taucs.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_taucs.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_taucs.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_taucs.NumericFactorization( ) );
EPETRA_CHK_ERR( A_taucs.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARDISO
} else if ( SparseSolver == PARDISO ) {
Teuchos::ParameterList ParamList ;
Amesos_Pardiso A_pardiso( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_pardiso.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_pardiso.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_pardiso.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_pardiso.NumericFactorization( ) );
EPETRA_CHK_ERR( A_pardiso.Solve( ) );
#endif
#ifdef HAVE_AMESOS_PARAKLETE
} else if ( SparseSolver == PARAKLETE ) {
Teuchos::ParameterList ParamList ;
Amesos_Paraklete A_paraklete( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_paraklete.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_paraklete.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_paraklete.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_paraklete.NumericFactorization( ) );
EPETRA_CHK_ERR( A_paraklete.Solve( ) );
#endif
#ifdef HAVE_AMESOS_MUMPS
} else if ( SparseSolver == MUMPS ) {
Teuchos::ParameterList ParamList ;
Amesos_Mumps A_mumps( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_mumps.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_mumps.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_mumps.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_mumps.NumericFactorization( ) );
EPETRA_CHK_ERR( A_mumps.Solve( ) );
#endif
#ifdef HAVE_AMESOS_SUPERLU
} else if ( SparseSolver == SUPERLU ) {
Teuchos::ParameterList ParamList ;
Amesos_Superlu A_superlu( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_superlu.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_superlu.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_superlu.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_superlu.NumericFactorization( ) );
EPETRA_CHK_ERR( A_superlu.Solve( ) );
#endif
#ifdef HAVE_AMESOS_LAPACK
} else if ( SparseSolver == LAPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Lapack A_lapack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_lapack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_lapack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_lapack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_lapack.NumericFactorization( ) );
EPETRA_CHK_ERR( A_lapack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_UMFPACK
} else if ( SparseSolver == UMFPACK ) {
Teuchos::ParameterList ParamList ;
Amesos_Umfpack A_umfpack( Problem ) ;
ParamList.set( "MaxProcs", -3 );
EPETRA_CHK_ERR( A_umfpack.SetParameters( ParamList ) );
EPETRA_CHK_ERR( A_umfpack.SetUseTranspose( transpose ) );
EPETRA_CHK_ERR( A_umfpack.SymbolicFactorization( ) );
EPETRA_CHK_ERR( A_umfpack.NumericFactorization( ) );
EPETRA_CHK_ERR( A_umfpack.Solve( ) );
#endif
#ifdef HAVE_AMESOS_KLU
} else if ( SparseSolver == KLU ) {
using namespace Teuchos;
Amesos_Time AT;
int setupTimePtr = -1, symTimePtr = -1, numTimePtr = -1, refacTimePtr = -1, solveTimePtr = -1;
AT.CreateTimer(Comm, 2);
AT.ResetTimer(0);
Teuchos::ParameterList ParamList ;
// ParamList.set("OutputLevel",2);
Amesos_Klu A_klu( Problem );
ParamList.set( "MaxProcs", -3 );
ParamList.set( "TrustMe", false );
// ParamList.set( "Refactorize", true );
EPETRA_CHK_ERR( A_klu.SetParameters( ParamList ) ) ;
EPETRA_CHK_ERR( A_klu.SetUseTranspose( transpose ) );
setupTimePtr = AT.AddTime("Setup", setupTimePtr, 0);
EPETRA_CHK_ERR( A_klu.SymbolicFactorization( ) );
symTimePtr = AT.AddTime("Symbolic", symTimePtr, 0);
EPETRA_CHK_ERR( A_klu.NumericFactorization( ) );
numTimePtr = AT.AddTime("Numeric", numTimePtr, 0);
EPETRA_CHK_ERR( A_klu.NumericFactorization( ) );
refacTimePtr = AT.AddTime("Refactor", refacTimePtr, 0);
// for ( int i=0; i<100000 ; i++ )
EPETRA_CHK_ERR( A_klu.Solve( ) );
solveTimePtr = AT.AddTime("Solve", solveTimePtr, 0);
double SetupTime = AT.GetTime(setupTimePtr);
double SymbolicTime = AT.GetTime(symTimePtr);
double NumericTime = AT.GetTime(numTimePtr);
double RefactorTime = AT.GetTime(refacTimePtr);
double SolveTime = AT.GetTime(solveTimePtr);
std::cout << __FILE__ << "::" << __LINE__ << " SetupTime = " << SetupTime << std::endl ;
std::cout << __FILE__ << "::" << __LINE__ << " SymbolicTime = " << SymbolicTime - SetupTime << std::endl ;
std::cout << __FILE__ << "::" << __LINE__ << " NumericTime = " << NumericTime - SymbolicTime<< std::endl ;
std::cout << __FILE__ << "::" << __LINE__ << " RefactorTime = " << RefactorTime - NumericTime << std::endl ;
std::cout << __FILE__ << "::" << __LINE__ << " SolveTime = " << SolveTime - RefactorTime << std::endl ;
#endif
} else {
SparseDirectTimingVars::log_file << "Solver not implemented yet" << std::endl ;
std::cerr << "\n\n#################### Requested solver not available on this platform ##################### ATS\n" << std::endl ;
std::cout << " SparseSolver = " << SparseSolver << std::endl ;
std::cerr << " SparseSolver = " << SparseSolver << std::endl ;
}
SparseDirectTimingVars::SS_Result.Set_Total_Time( TotalTime.ElapsedTime() );
} // end for (int i=0; i<special; i++ )
//
// Compute the error = norm(xcomp - xexact )
//
double error;
passresid->Update(1.0, *passx, -1.0, *passxexact, 0.0);
passresid->Norm2(&error);
SparseDirectTimingVars::SS_Result.Set_Error(error) ;
// passxexact->Norm2(&error ) ;
// passx->Norm2(&error ) ;
//
// Compute the residual = norm(Ax - b)
//
double residual ;
passA->Multiply( transpose, *passx, *passtmp);
passresid->Update(1.0, *passtmp, -1.0, *passb, 0.0);
// passresid->Update(1.0, *passtmp, -1.0, CopyB, 0.0);
passresid->Norm2(&residual);
SparseDirectTimingVars::SS_Result.Set_Residual(residual) ;
double bnorm;
passb->Norm2( &bnorm ) ;
SparseDirectTimingVars::SS_Result.Set_Bnorm(bnorm) ;
double xnorm;
passx->Norm2( &xnorm ) ;
SparseDirectTimingVars::SS_Result.Set_Xnorm(xnorm) ;
delete readA;
delete readx;
delete readb;
delete readxexact;
delete readMap;
delete map_;
Comm.Barrier();
return 0;
}
